Linguistics A path through The Default Library ======================================== How the world's seven thousand languages work, how they change, how children learn them and how machines model them. Step 1. Phonetics and Phonology: The Sounds You Never Knew You Were Choosing ---------------------------------------------------------------------------- Why: How speech sounds are made, and the patterns each language makes of them. Every language uses a few dozen sounds out of the hundred-odd a human mouth can make, and its speakers stop hearing the difference between the ones their language ignores. Phonetics measures the sounds; phonology finds the system behind them. Hold your hand in front of your mouth and say "pin", then "spin". In "pin" you feel a small puff of air on your palm. In "spin" you feel almost none. Those are two different sounds, as different in your mouth as a "p" and a "b" are to someone else. Yet you have probably never noticed, and an English speaker would call both of them "the letter p". A Hindi speaker would not. In Hindi the puffed and unpuffed versions separate words: pal means "moment" and phal (with the puff) means "fruit". The mouth makes the same two sounds in both languages. What differs is which differences the language has decided to count. That gap, between what the mouth does and what the mind keeps track of, is the whole subject, and it is why linguistics courses start here. The course in one box Who takes itLinguistics majors, in year 1, usually as two linked courses (phonetics, then phonology) or one combined one. The Linguistic Society of America's guidance for the undergraduate major, the UK's QAA benchmark for linguistics and the introductory sequence at MIT (Course 24) all put it at the start. Speech-language pathologists, language teachers and speech-recognition engineers take it too. What it is forDescribing every speech sound in any language by how it is made, and finding the rules that decide which sounds a language uses and how it arranges them. Comes afterNothing: it is the entry point to the major. Leads toMorphology (36/22), Historical Linguistics (36/25), Sociolinguistics (36/26), Language Acquisition (36/27), Psycholinguistics and the Brain (36/28) and Computational Linguistics (36/30). It sits beside Languages of the World (36/29). What this piece gives youYou will not be able to transcribe a new language by ear. You will be able to read a place-and-manner chart, see why a vowel is a point on a map, tell a contrast from a variation, and know why tone and clicks are not exotic. Easier first reads: Phonetics and the IPA (26/01) and Waves and Sound (13/13). The sound tools here are synthesised in your browser; nothing is a recording, and every one also works silently. Place, manner and voicing: every consonant is three choices A consonant is a gesture that blocks or narrows the air on its way out. Phonetics describes it with three answers. Where is the narrowing (place)? Lips for [p], tongue tip behind the teeth for [t], the back of the tongue against the soft palate for [k]. How does the air get past (manner)? Stopped completely and released (a plosive: [p t k]), forced through a slit with hiss (a fricative: [s f]), sent through the nose (a nasal: [m n]), and so on. And is the larynx buzzing (voiced: [z b m]) or silent (voiceless: [s p])? Those three answers are enough to name nearly every consonant in the world: [s] is a voiceless alveolar fricative, and [z] is the same with the buzz switched on. Put your fingertips on your throat and hold "sss", then "zzz". The International Phonetic Alphabet, founded by language teachers in Paris in 1886, is a chart of this idea: columns for place, rows for manner, voiceless and voiced symbols side by side. Interactive · A piece of the IPA chart Tap a symbol. Blue underline: voiceless. Red underline: voiced. The chart scrolls sideways on a narrow screen. Stop sound Sound is a rough synthesis (noise and buzz shaped by filters), enough to hear fricatives, nasals and stop-plus-vowel; trills, taps and laterals are shown without sound. It is not a recording of a speaker. This is a subset: the full chart also has pharyngeal, retroflex and many more. Without JavaScript: [p] voiceless bilabial plosive; [b] voiced bilabial plosive; [s] voiceless alveolar fricative; [z] voiced alveolar fricative; [m] voiced bilabial nasal. What a graduate remembers is that the chart has gaps with reasons. English has no voiceless nasal phonemes (a silent [m]), though languages such as Burmese do. A bilabial fricative [ɸ] is rare in English but is the "f" in Japanese fu. Two places can be used at once: the English "w" is made with the lips and the back of the tongue. And note what the chart does not contain: letters. Spelling is a different system, which is why English ough can be pronounced six ways. Vowels are a map, and formants are the coordinates Vowels have no narrowing at all, so "place" is not the right question. Instead, a vowel is a shape of the mouth: how high the tongue is (and so how open the jaw), how far forward it is, and whether the lips are rounded. Say "ee, ay, ah" slowly and feel the jaw drop. Say "ee" then "oo" and feel the tongue slide back and the lips round. The shape matters because the mouth is a resonating tube. The buzz from your larynx contains many frequencies, and the tube boosts a few of them. Those boosted bands are formants. The lowest, F1, goes up as the jaw opens: "ee" has a low F1 and "ah" a high one. The second, F2, goes up as the tongue moves forward: "ee" has a high F2 and "oo" a low one. Plot F1 downward and F2 right-to-left and you get the familiar vowel chart, drawn from acoustics instead of from tongue positions. This is one of the cleaner results in the field: the old tongue-position chart and the formant plot roughly coincide. F1 ↑ as the mouth opens  ·  F2 ↑ as the tongue moves forwardThe two numbers locate a vowel. Lip rounding lowers F2 a little more. Interactive · Drive the vowel space Tap or drag on the map, or use the sliders. "Hold" buzzes a vowel until you stop it. Formant values for the 16 labelled vowels are averages for a male voice (Catford 2001). F1 (jaw: closed to open) F2 (tongue: back to front) Hold soundStop sound Synthesised with a 120 Hz buzz passed through three resonant filters. Real vowels vary by speaker: a child's formants are higher than a man's, yet we hear the same vowel, because the brain compares the pattern, not the numbers. Without JavaScript: [i] has F1 about 240 Hz and F2 about 2400 Hz; [a] has about 850 and 1610; [u] has about 250 and 595. This is also why the symbols mean less than they look. Your "ee" and mine differ in the numbers; the IPA labels a region of the map, not a point. And a vowel in your language is judged against its neighbours: Spanish has five vowels and spreads them wide; a language with fifteen must crowd them. Phonemes: the sounds a language counts Now the turn from phonetics to phonology. The mouth can make a near-infinite number of sounds, but each language carves them into a small set of phonemes: categories that change meaning. Two sounds are different phonemes if swapping one for the other can turn one word into another. The test is the minimal pair: pat and bat, which differ only in voicing, prove that English /p/ and /b/ are different phonemes. (Slashes mean phoneme; square brackets mean the actual sound.) If two sounds never swap words, they are allophones of one phoneme. The puffed [pʰ] of "pin" and the unpuffed [p] of "spin" are allophones of English /p/: the puffed one comes at the start of a stressed syllable, the plain one after /s/. Their distributions don't overlap (complementary distribution), so the language never needed to use them to tell words apart, and speakers never learned to hear the difference. The same pair of sounds is two phonemes in Hindi, Thai and Korean. Interactive · Contrast or variation? Pick a language. Look at the words. Decide whether the two sounds are separate phonemes, then check. One phoneme (variation)Two phonemes (contrast) Without JavaScript: Hindi pal [pəl] "moment" and phal [pʰəl] "fruit" are a minimal pair, so [p] and [pʰ] are two phonemes. In English they are allophones of /p/. In Spanish [d] and [ð] are one phoneme; in Japanese [t], [ts] and [tɕ] are one. This is why a learner's "accent" is not carelessness. A Japanese speaker, whose language has a single liquid phoneme, has no trained category for the English /l/-/r/ split; a Spanish speaker hears [d] and [ð] as one sound and will not notice that they have been using both all their life. Babies start with every contrast open and, by about a year, tune toward the ones their language uses. Rules over features: why "cats" and "dogs" end differently Phonology's second discovery is that sounds change according to rules that refer to features, not to individual sounds. Take the English plural. Spelling shows an "s" for all three: cats, dogs, buses. In speech they are different: cats [kæts], dogs [dɔɡz], buses [ˈbʌsɪz]. Nobody teaches children this, yet they apply it to words they have never heard. In Jean Berko's 1958 "wug test", children shown a picture of one invented creature and then two supplied "wugz", with the voiced ending, to a word that did not exist. The rule has two steps, and both are about features. After a sibilant (a hissing sound like [s z ʃ ʒ tʃ dʒ]) add the extra vowel, [ɪz], because otherwise the two hisses would run together. Otherwise copy the buzz of the last sound: voiceless stays voiceless ([s]), voiced stays voiced ([z]). The plural suffix quietly agrees with the sound before it. Other languages do similar things with bigger consequences. In German, voiced stops lose their voicing at the end of a word, so Rad "wheel" and Rat "advice" are both pronounced [ʁaːt]; the plurals Räder and Räte bring back the difference. Interactive · Make the plural Pick the last sound of a singular noun. The page applies the two-step rule. Without JavaScript: cat gives [s] (voiceless ending), dog gives [z] (voiced), bus gives [ɪz] (sibilant). Tone: when pitch spells the word English uses pitch for questions and moods, but not to tell words apart. In a tone language the pitch contour is part of the word, as much as its consonants. WALS, the World Atlas of Language Structures, sorts 527 languages in Maddieson's chapter on tone: 307 have no tones, 132 have a simple system (usually two contrasts, high and low), and 88 have a complex one. So well over a third of the sample uses pitch for words, and that is not just East Asia: Yoruba and Igbo in West Africa and Mixtec in Mexico are tonal. Standard Mandarin has four tones. The syllable ma is "mother", "hemp", "horse" or "scold" depending on whether the pitch stays high, rises, dips and rises, or falls sharply. Cantonese counts six, in the common Jyutping description: the syllable si alone covers "poem", "history", "try", "time", "market" and "matter". Linguists write contours with Chao's five-level number system, 5 being the top of the speaker's range, so Mandarin's tones are 55, 35, 214 and 51. Interactive · The tone lab Pitch levels 1 to 5 map to 110 to 220 Hz in this synthesis. Real speakers' tones also vary in length and voice quality, and tones change in running speech. Contours are the citation forms described for Standard Mandarin and Cantonese. Without JavaScript: Mandarin ma: 55 mother, 35 hemp, 214 horse, 51 scold. Cantonese si: 55 poem, 25 history, 33 try, 21 time, 23 market, 22 matter. Clicks, and how big a sound inventory can be Languages differ in how many sounds they count. WALS (Maddieson) puts 563 languages in five size bands of consonant inventories: 89 are small (6 to 14 consonants), 122 moderately small, 201 average (about 19 to 25), 94 moderately large and 57 large (34 or more). The extremes are 6 consonants in Rotokas, spoken on Bougainville Island in Papua New Guinea, and up to 122 in the Khoisan language !Xóõ, now usually called Taa, spoken mainly in Botswana by about 2,500 people. Much of Taa's size comes from clicks, which are made by trapping a pocket of air with the tongue and releasing it with a suction pop. Clicks are phonemes in many languages of southern Africa (Zulu, Xhosa, Nama, Taa) and in two languages of Tanzania, Hadza and Sandawe. They are not odd mouth-noises. Xhosa, with several million speakers, writes three click places as c, q and x, so ucango "door", uqeqesho "training" and the first sound of isiXhosa itself all start with clicks. And you already make clicks: a disapproving "tsk tsk" is the dental click, a horse-clop is the lateral click, and a kiss is the bilabial one. English just doesn't use them in words. Interactive · Five clicks Stop sound A very rough synthesis of the short burst each click makes. Real clicks are produced with a second articulation (voicing, nasality, aspiration), which is how Xhosa gets 15 to 21 of them from three places. Without JavaScript: dental ǀ is "tsk"; lateral ǁ is the horse-clop sound; bilabial ʘ is a kiss; alveolar ǃ and palatal ǂ are popping clicks. A last fact about inventories: they are not the whole story. Two languages with the same sounds can have very different phonotactics, the rules for how sounds may combine. English allows the cluster "str" at the start of a word; Japanese basically allows consonant, vowel, consonant, vowel, and a Japanese speaker will insert vowels into English words, turning "street" into sutoriito. Phonology covers inventory, combination and change together. What a graduate just sees Every consonant is a place, a manner and a voicing; a gap in the chart usually has a reason. A vowel is a point on a map; F1 follows jaw opening, F2 follows tongue frontness. The unit that matters is the phoneme: a sound counts only if it can change a word. Two sounds that never appear in the same place are usually one phoneme, and the speaker can't hear the split. Rules refer to features (voicing, sibilance), which is why the plural agrees with the sound before it, and why it applies to made-up words. Spelling is not sound; a foreign accent is a trained set of categories, not carelessness. Tone and clicks are ordinary equipment for well over a third of the world's languages, and the size of sound inventories ranges from about 20 to over 100. Where it shows up next Morphology (36/22): the English plural has three shapes; Turkish vowel harmony is a phonological rule that reaches into word-building. Historical Linguistics (36/25): sound change is regular, which is how related languages are proved to be related. Sociolinguistics (36/26): accents are patterns of allophones and vowel shifts, and they carry social meaning. Language Acquisition (36/27) and Psycholinguistics (36/28): infants tuning to their language's contrasts; how listeners recognise words in noise. In the world: speech recognition and text-to-speech engines, speech therapy, dictionary pronunciations, and the design of writing systems for languages that have none. If you want to go deeper 📘 Textbook (phonetics). Peter Ladefoged and Keith Johnson, A Course in Phonetics (7th edition, 2015): the standard first course, with practical transcription. 📘 Textbook (phonology). David Odden, Introducing Phonology (2nd edition, Cambridge University Press, 2013), which teaches the problem-solving method with data from many languages. 📘 Gentler. Peter Ladefoged and Sandra Disner, Vowels and Consonants, an accessible tour of the world's sounds. 🌐 Free text. Essentials of Linguistics, 2nd edition (Anderson and colleagues, McMaster University), a free open textbook with phonetics and phonology chapters: ecampusontario.pressbooks.pub/essentialsoflinguistics2. 🌐 Hear the real thing. The interactive chart at ipachart.com plays recorded speakers for each symbol. 🌐 Free course. MIT OpenCourseWare 24.900 Introduction to Linguistics (Spring 2022). The takeaway Speech sounds are gestures you can name by place, manner and voicing, and vowels are points on a formant map. But the mind does not keep every sound: each language keeps a small set of phonemes, groups the rest as allophones, and shapes them with rules about features. That is why accents persist, why spelling misleads, and why a click or a tone is no stranger than a "p". Try this: say "pin" and "spin" with a tissue in front of your mouth, then say "kit" and "skit". Then listen for what your own language does at the ends of words. Sources & Further Reading Sources. Vowel formants: averages for a male voice from Catford (2001), as tabulated in Wikipedia's "Formant" article. Tone: WALS Online chapter 13 (Maddieson), table counts 307, 132 and 88 of 527; the text of the chapter gives slightly different totals in places. Consonant inventories: WALS chapter 1 (Maddieson), 563 languages. Rotokas (6 consonants in Central Rotokas, Bougainville) and Taa (about 2,500 speakers in 2011; consonant counts depend on the analysis, at least 58 in one and 87 in another; the WALS maximum of 122 is for !Xóõ) from Wikipedia's articles on those languages. Xhosa clicks and example words from Wikipedia's "Xhosa language". Berko, "The child's learning of English morphology", Word 14 (1958). Mandarin tone numbers and Cantonese Jyutping tones as described in Wikipedia's articles on Standard Chinese phonology and Cantonese phonology; Chao tone numbers. German final devoicing is textbook. IPA founded 1886: see Phonetics and the IPA (26/01). The numbers on this page (formants, pitch mapping, plural rule) were run in node. Example words and transcriptions are standard but simplified; real speakers vary. 📘 Ladefoged and Johnson, A Course in Phonetics 📘 Odden, Introducing Phonology 🌐 ipachart.com Step 2. Morphology: How Words Are Built --------------------------------------- Why: How words are built. English builds "unbreakableness" from four pieces and Turkish builds "from your houses" from four more, in a single word. Morphology is the study of those pieces, and of the surprisingly different ways languages put them together. The word unbreakableness has never been in a dictionary, because nobody needs to look it up. You can see how it works: break is the root, -able turns it into "capable of being broken", un- reverses that, and -ness makes the result a thing you can talk about. Four pieces, each with its own job, snapped together by a rule you absorbed as a child and never wrote down. Now take the Turkish word evlerinizden. It is one word, and it means "from your houses" (you plural). It is also four pieces: ev "house", -ler plural, -iniz "your", -den "from". Chinese would use separate words for most of that; Inuktitut and Yup'ik would pack in a whole sentence. How much goes inside the word, and how it is bolted on, is what morphology describes. The course in one box Who takes itLinguistics majors, usually in year 1 or 2, after a first course in sounds. Programmes that follow the Linguistic Society of America's guidance, the UK QAA benchmark for linguistics and MIT's Course 24 all include it. Language teachers and computational linguists use it constantly. What it is forTaking words apart into their smallest meaningful pieces and describing the rules for putting them together. Comes afterPhonetics and Phonology (36/21), because the shape of a piece often depends on the sound next to it. Leads toSyntax (36/23), which is where pieces become sentences; and Historical Linguistics (36/25), which uses endings to prove languages are related. It sits beside Languages of the World (36/29). What this piece gives youYou will not be able to analyse a language you have never met. You will be able to cut a glossed word into morphemes, see how Arabic builds words on a skeleton, build Turkish forms yourself, and say why "isolating, agglutinative, fusional, polysynthetic" are tendencies rather than boxes. Easier first read: Where English Words Come From (26/03). Morphemes: the smallest pieces that carry meaning A morpheme is the smallest unit that pairs a form with a meaning or a grammatical job. It is not the same as a syllable (the word cats has one syllable and two morphemes) and not the same as a word (unbreakableness has four). Some morphemes can stand alone and are called free (break); some must attach and are bound (-able, -ness). The core of a word is its root; what attaches is an affix. To compare languages, linguists write examples in a standard way, the Leipzig Glossing Rules: the word in the language with hyphens at the morpheme boundaries, a line below it giving a gloss for each piece (small capitals or abbreviations for grammatical ones, such as PL for plural, 1SG for "I"), and a free translation. The tool below lets you cut nine words, from every inhabited continent, along their seams. Interactive · Cut the word Tap a piece to see what it is. Glosses follow the Leipzig Glossing Rules; abbreviations: PL plural, POSS possessive, ABL "from", FUT future, NEG negative, CAUS causative, FV final vowel, PST.HAB past habitual, S subject, SG singular. Analyses simplified; the Yup'ik, Nahuatl, Chukchi and Tiwi examples are as published by their cited authors (see Sources). Without JavaScript: Turkish ev-ler-iniz-den, house-PL-2PL.POSS-ABL, "from your houses". Swahili ni-ta-ku-pend-a, 1SG-FUT-2SG.OBJ-love-FV, "I will love you". Roots, affixes and patterns: three ways to build Most people picture a word as a root with things stuck to its ends, and that is the commonest method: suffixing (English -ness, Turkish -ler) is more frequent in the world's languages than prefixing (Swahili ni- "I", English un-). But there are other options. Infixing puts the piece inside: in Tagalog (the Philippines), sulat "write" becomes sumulat, with -um- marking the actor. Circumfixing wraps both sides: German spielen "play" gives ge-spiel-t "played", and Indonesian besar "big" gives ke-besar-an "greatness". Reduplication repeats the root: Indonesian buku "book" gives buku-buku "books". The most striking method belongs to the Semitic languages, including Arabic and Hebrew. There, a word's root is not a string of letters in order but a set of three consonants carrying a core meaning, and the vowels belong to separate patterns that are poured over the skeleton. The root k-t-b is about writing; put it into one pattern and you get kataba "he wrote", into another kitāb "book", into another maktaba "library". The same patterns work on other roots, so madrasa "school" and maktaba "library" share a place-of pattern, built on d-r-s "study" and k-t-b "write". Interactive · Root and pattern Consonants of the root are in red, the pattern's own letters in blue. Only attested common words are listed; not every root fills every pattern. Arabic is shown in standard transliteration. Without JavaScript: k-t-b gives kataba "he wrote", kitāb "book", kātib "writer", maktab "office", maktaba "library"; d-r-s gives darasa "he studied", madrasa "school", mudarris "teacher". Inflection and derivation: changing the form or the job Affixes come in two kinds, and the difference runs through the whole field. Derivation makes a new word: happy to happiness changes an adjective to a noun and would get its own dictionary entry; teach to teacher names a new kind of thing. Inflection gives a form of the same word that grammar demands: dog to dogs, walk to walked. Inflection is usually regular, doesn't change the word class, and sits on the outside. In English the order is nearly always root, then derivational pieces, then inflection: nation-al-iz-ation-s, never the other way round. Words also have structure, not just a sequence. Unlockable is ambiguous. If un- attaches to lock first, you get unlock, and then -able makes it "able to be unlocked" (a door you can open). If -able attaches to lock first, you get "lockable", and un- reverses it: "not able to be locked" (a door that will not lock). Same letters, two trees, two meanings. Interactive · Two trees for one word Without JavaScript: [un-[lock-able]] means "not lockable"; [[un-lock]-able] means "able to be unlocked". Allomorphs and irregulars: one meaning, several shapes A morpheme can take different shapes in different places; those shapes are its allomorphs. English plural -s is pronounced [s] in cats, [z] in dogs and [ɪz] in buses, chosen by the last sound of the word, as the plural tool in Phonetics and Phonology (36/21) shows. The past tense does the same: walked [t], jogged [d], wanted [ɪd]. Children apply these to words they have never heard. In Jean Berko's 1958 experiment they were shown an invented creature, "a wug", and then two of them, and said "wugz". Then there are the exceptions: sang, geese, went. Irregulars are mostly fossils. Old English had a large class of verbs that changed their vowel to mark the past (sing, sang, sung), and over centuries most were regularised. A famous 2007 study by Erez Lieberman and colleagues followed 177 Old English irregular verbs: 145 were still irregular in Middle English, and 98 are irregular today. The striking result is which ones survive: the regularisation rate falls with the square root of how often a verb is used. A verb a hundred times as common takes about ten times as long to lose its irregular form. Frequent verbs such as be, have and go are the most irregular words in English precisely because everyone hears them so often that each form is memorised. Interactive · How long does an irregular last? Verb A is this many times more frequent than verb B The scaling law only: the half-life of an irregular form grows as the square root of its frequency (Lieberman et al. 2007). The real half-lives in the paper (up to thousands of years for the commonest verbs) depend on measured frequencies not reproduced here. Without JavaScript: a verb 100 times more frequent keeps its irregular form about 10 times longer; 10,000 times more frequent, about 100 times. How the mind handles all this is a live debate. One camp, with Steven Pinker and Alan Prince, argues for words and rules: regulars come from a rule, irregulars are stored in memory. The other, going back to David Rumelhart and James McClelland's 1986 network model, holds that a single pattern-learning system handles both. Both sides agree on the data; they disagree on the machinery, and the language models of today (36/30) are a continuation of that argument. Agglutination: building Turkish by hand In an agglutinative language, such as Turkish, Finnish, Swahili, Quechua or Japanese, each affix carries one meaning and the pieces line up in a fixed order like train carriages. That makes the system easy to compute, so you can build the forms yourself. Turkish has one more feature: vowel harmony. The vowel in each suffix copies features from the last vowel before it, front or back and rounded or not, so that -ler becomes -lar after back vowels (okul-lar "schools") and -ler after front ones (ev-ler "houses"). Harmony is a sound rule (36/21) reaching into word-building. Pick a noun, a number, a possessor and a case. The page assembles the word with the harmony and consonant rules. A single noun stem here gives 2 × 6 × 5 = 60 forms; with the seven stems that is 420 words, all from a handful of pieces. Interactive · Build a Turkish noun Noun Number Possessor Case Rules built in: 2-way harmony (a/e), 4-way harmony (ı/i/u/ü), a buffer consonant (y, s, n) after vowels or after a third-person possessor, and d/t after voiceless consonants. Stems that change a final consonant before a vowel (such as kitap to kitabı) are left out. Without JavaScript: ev + ler + iniz + den gives evlerinizden "from your houses"; okul + lar + ımız + da gives okullarımızda "in our schools". Four ways to pack meaning, and why they blur Nineteenth-century scholars sorted languages into types, and the four labels survive. Isolating (or analytic) languages, such as Mandarin and Vietnamese, use mostly one morpheme per word and leave grammar to word order and little function words: tā mǎi le sān běn shū, "he bought three books", is six words and six morphemes. Agglutinative languages stack single-meaning affixes, as above. Fusional languages pack several meanings into one inseparable ending: in Latin amō, the single -ō means "I", "now", "is a statement" and "active" all at once; you cannot cut it into four. And polysynthetic languages build whole sentences into one word. Yup'ik of Alaska can say tuntussuqatarniksaitengqiggtuq, "he had not yet said again that he was going to hunt reindeer". Nahuatl in Mexico, Chukchi in Siberia and Tiwi in northern Australia do something similar. Linguists use these as tendencies, not boxes. Joseph Greenberg proposed in 1960 to count morphemes per word in a sample text (an "index of synthesis", with under 2 called analytic, 2 to 3 synthetic and above 3 polysynthetic, as it is usually reported). Critics point out that the labels bundle several separate dimensions, that most languages mix types (English is usually filed under fusional but is quite analytic), and that the 19th-century scheme came with a bias: some scholars ranked agglutinative or inflecting languages as more "advanced". That ranking is discredited; every language builds all the meaning its speakers need. WALS now measures the dimensions separately: how much the pieces fuse, how many meanings an ending carries, and how many places a piece can occur. Interactive · Which type is it? The numbers are computed from the gloss shown (one example sentence is not an index of the whole language). "Meanings per ending" counts grammatical features in each affix gloss. Without JavaScript: Mandarin tā mǎi le sān běn shū: 6 words, 6 morphemes (isolating). Latin amō: -ō carries 4 meanings (fusional). Yup'ik tuntussuqatarniksaitengqiggtuq: one word, 7 morphemes (polysynthetic). What a graduate just sees A word is not the unit of meaning; the morpheme is, and a word can be one or fifteen of them. Derivation makes new words; inflection makes the form grammar asks for, usually outside the derivation. Structure matters: unlockable has two trees, and so two meanings. One morpheme can have several shapes (allomorphs), chosen by sounds nearby. Irregular words survive by being common: usage frequency protects them. Languages can build by prefix, suffix, infix, circumfix, repetition or root-and-pattern; none is more advanced than another. "Isolating, agglutinative, fusional, polysynthetic" describe tendencies and dimensions, not boxes. Where it shows up next Syntax (36/23): in a language with rich endings, word order can be freer, because the endings mark who did what. Languages of the World (36/29): typology across the globe; WALS features on fusion and exponence. Language Acquisition (36/27): children's over-regularisations ("goed", "foots") show them learning the rule. Computational Linguistics (36/30): stemmers, tokenisers and language models have to decide where words break, and agglutinative languages like Finnish and Turkish make that hard. In the world: spelling checkers, search engines that match "run" and "running", dictionaries, and machine translation into languages with many more word forms than English. If you want to go deeper 📘 Textbook. Martin Haspelmath and Andrea Sims, Understanding Morphology (2nd edition, Routledge), which draws its examples from languages worldwide. 📘 Textbook. Mark Aronoff and Kirsten Fudeman, What Is Morphology? (Wiley-Blackwell, several editions), a shorter first course with exercises. 📘 Gentler. Steven Pinker, Words and Rules (1999), on the regular and irregular debate, written for general readers. 🌐 Free text. Essentials of Linguistics, 2nd edition (McMaster University), free online, with a morphology chapter: ecampusontario.pressbooks.pub/essentialsoflinguistics2. 🌐 Free course. MIT OpenCourseWare 24.900 Introduction to Linguistics (Spring 2022), which covers words and their structure. The takeaway Words are built from morphemes by rules. Languages differ in how many pieces go in a word, whether each piece carries one meaning or many, and how the pieces are joined, from Turkish carriages to Arabic skeletons to Yup'ik sentence-words. But the types are tendencies, and no way of building is a step up or down. Try this: pick a long word you met today and cut it into morphemes. Then check each piece: is it a root, a derivational affix or an inflection, and does it have an allomorph? Sources & Further Reading Sources. Leipzig Glossing Rules (Max Planck Institute for Evolutionary Anthropology and the University of Leipzig). Yup'ik and other polysynthetic examples: Payne (1997) for Yup'ik; Skorik (1961) for Chukchi; Suárez (1983) for Nahuatl; Blake (2006) for Tiwi, as reproduced in Wikipedia's "Polysynthetic language". Morphological typology and its criticisms, including WALS's separate measures: Wikipedia's "Morphological typology", with Greenberg (1960), "A quantitative approach to the morphological typology of language", as described in secondary sources (thresholds 2 and 3 as reported there). Berko, Word 14 (1958). Lieberman, Michel, Jackson, Tang and Nowak, "Quantifying the evolutionary dynamics of language", Nature 449 (2007): 177 Old English irregular verbs, 145 in Middle English, 98 today, half-life proportional to the square root of frequency. Rumelhart and McClelland (1986); Pinker and Prince (1988); Pinker, Words and Rules (1999). Arabic root-and-pattern examples (k-t-b, d-r-s) are standard in Arabic grammars. Turkish suffix rules are from standard Turkish grammars; the forms above were generated by the page's code and run in node (for example evlerinizden). Tagalog -um-, German ge-...-t and Indonesian ke-...-an and reduplication are textbook examples. The Swahili, Quechua (Cusco), Indonesian and Mandarin examples are standard textbook analyses, simplified. 📘 Haspelmath and Sims, Understanding Morphology 📘 Pinker, Words and Rules 🌐 Essentials of Linguistics Step 3. Languages of the World: Seven Thousand Answers to the Same Problem -------------------------------------------------------------------------- Why: Families, types and the diversity of the world's languages. Every human community solved the problem of talking to each other, and no two solved it the same way. Typology is the science of how many different solutions there are, which ones go together, and which ones are about to vanish. Count the languages in a village in Papua New Guinea and you may need both hands for the households. Ask how many languages the whole planet has and you get an answer that depends on who you ask: about seven thousand, give or take hundreds. The two are connected. Most of the world's languages are small, most of them live in a few crowded regions, and a handful of giants carry most of the world's speakers. This is the subject that lets you zoom out. A first course looks at the sounds of one language, then its words, then its sentences. Languages of the World (sometimes called linguistic typology and diversity) steps back to look at all of them at once: how they are counted, how they are related, how they differ in grammar, which are in danger, and how they are written. The course in one box Who takes itLinguistics majors, in year 1 or 2; anthropology and area-studies students often take a "languages of the world" course too. It is a standard component of undergraduate programmes that follow the Linguistic Society of America's guidance or the UK QAA benchmark, and shows up in MIT's Course 24 introduction as units on language change and endangerment. What it is forGiving you a map of the whole field of human languages: the families, the range of structures, the endangered ones and the writing systems, with data you can check. Comes afterMorphology (36/22). Leads toHistorical Linguistics (36/25), which proves families; Sociolinguistics (36/26), on language policy and revival; and Syntax (36/23), where word-order types are explained. What this piece gives youYou will not be able to classify a language you have not met. You will see why "how many languages" has no single answer, read a WALS count, recognise the four revivals, and build a Korean syllable from its parts. Easier first reads: The Family Tree of Languages (26/04) and How Language Changes (26/05). How many languages are there? It depends what you count Ethnologue, the reference most people quote, counts 7,170 living languages in its 29th edition (2026), reported as the highest number of living languages it has listed; its 27th edition (2024) had 7,164. Glottolog, an academic catalogue, counts 7,674 spoken first languages in version 5.3 (March 2026), grouped into 246 families and 183 isolates. Neither is wrong. They draw the line between "a language" and "a dialect" in different places, and the line is genuinely hard to draw. The usual test is mutual intelligibility: if speakers understand each other without training, they speak the same language. But intelligibility comes in degrees, can be one-way, and shifts with exposure. Politics also matters. Swedish, Danish and Norwegian are largely mutually intelligible but are counted as three languages; the "dialects" of Chinese often are not intelligible to each other but are counted as one. A saying quoted by the Yiddish scholar Max Weinreich in 1945 puts it as "a language is a dialect with an army and a navy". He was repeating a remark made to him, but the point stands: the label is social as much as linguistic. Interactive · One chain of villages, many counts Twelve invented villages along a valley. Neighbours understand each other well; understanding fades with distance. Set how much understanding you demand before two villages count as one language. Intelligibility required to count as "the same language" Illustrative data: intelligibility between villages is invented as 100% minus 12 points for each step apart. Real intelligibility is uneven and often one-way. Without JavaScript: with the villages above, demanding 50% understanding gives 3 languages; demanding 95% gives 12; demanding almost nothing gives 1. The same chain gives one language, three languages or twelve, depending on the threshold, and nothing about the villages changed. That is a fair picture of the real disagreement. Families: trees, isolates and a lopsided world Languages are grouped by common ancestry into families, shown by the regular sound correspondences you will study in Historical Linguistics (36/25). English belongs to Indo-European, with Hindi, Persian, Spanish, Russian and Greek. Swahili belongs to Atlantic–Congo (the core of the Niger–Congo family), as do Yoruba and Zulu. Mandarin, Burmese and Tibetan are Sino-Tibetan. Tagalog, Malagasy, Māori and Hawaiian are Austronesian, a family that spread from Taiwan across the Pacific and Indian Oceans. Arabic, Hebrew and Hausa are Afroasiatic. Many languages of New Guinea are Trans–New Guinea. The world is lopsided in two ways. First, the biggest families by number of languages are not the biggest by speakers. Second, some languages belong to no known family at all. Linguists call these isolates: Basque in Europe, Burushaski in Pakistan, Hadza in Tanzania and Zuni in New Mexico are the standard examples, though a language's status as an isolate means only that no relationship has been shown, and some get reclassified. Interactive · Languages versus speakers Language counts from Glottolog 5.0 and speaker numbers from Ethnologue (2019), as tabulated in Wikipedia's "List of language families"; the two sources date differently and classify differently (for instance "Atlantic–Congo" is printed there in place of Niger–Congo). Treat the figures as rough. Without JavaScript: Indo-European has 586 languages and about 3.2 billion speakers; Austronesian has 1,274 languages and about 326 million; Trans–New Guinea has 376 languages and about 3.7 million. Indo-European is the giant of speakers, because a few of its languages spread through empires and migrations, while Trans–New Guinea and Austronesian have many languages and few speakers each. Where did Indo-European start? The long-running contest is between a steppe homeland north of the Black and Caspian seas about 6,000 years ago and an Anatolian one, thousands of years earlier. A 2023 Science paper by Paul Heggarty and colleagues, using vocabulary data from 161 languages, argued for a hybrid, with a root south of the Caucasus about 8,100 years ago. Other specialists dispute its methods. It is a live question. Word order: counting a feature across a thousand languages Typology asks which features go together. The great reference is the World Atlas of Language Structures (WALS), which maps a feature across hundreds of languages. For basic word order, Matthew Dryer's chapter sorts 1,376 languages by how the subject (S), verb (V) and object (O) are usually arranged in a plain statement. The result is not an even spread. Interactive · Six word orders, one sentence idea Counts: WALS chapter 81 (Dryer), 1,376 languages. Examples are glossed in the Leipzig style; Hixkaryana and Malagasy sentences are as published in their language articles (see Sources). "S" subject, "V" verb, "O" object. Without JavaScript: SOV 564 languages, SVO 488, VSO 95, VOS 25, OVS 11, OSV 4, no dominant order 189 (WALS, Dryer). Two orders, SOV (Japanese, Turkish, Hindi, Quechua) and SVO (English, Swahili, Mandarin), hold nearly nine in ten of the languages that have a dominant order (1,052 of 1,187). The rarer ones are real languages with real speakers: Welsh is VSO, Malagasy is VOS, Hixkaryana, spoken by about 600 people on the Nhamundá River in Brazil, is OVS. And the choice is not independent of other choices. Languages that put the verb after the object tend to use postpositions (words after the noun, like Japanese Tōkyō ni, "Tokyo to"), and languages that put the verb before the object tend to use prepositions (English "to Tokyo"). Interactive · Do the choices go together? WALS chapter 95 (Dryer): 1,142 languages, of which 158 do not fit these four combinations. Shares are computed from the counts. Without JavaScript: among OV languages, 472 have postpositions and 14 prepositions; among VO languages, 456 have prepositions and 42 postpositions. Typologists call this tendency harmony, and it was a headline result of Joseph Greenberg's 1963 survey of word-order universals. Why it exists is argued over: functional explanations point to ease of processing (keeping the "head" word on the same side), others to how languages change over time. A single tendency, with exceptions, is what a typologist means by "universal". Hold on to the idea that a feature counted over many languages shows what is common and what is rare, which is not the same as what is possible. Endangered languages: why transmission is the whole story Languages die when speakers stop passing them on, usually because of pressure to use a more powerful language in school, work and the market. UNESCO's Atlas of the World's Languages in Danger (third edition, 2010) catalogued 2,473 languages by degree of danger, from "vulnerable" to "critically endangered". Estimates of the future range widely; some linguists have projected that between half and nine-tenths of today's languages will be severely endangered or gone by 2100. That range is a projection, not a measurement, and some scholars think the worst figures overstate it. What is not disputed is the mechanism. Joshua Fishman, in 1991, built his scale of language shift around one test: do children grow up speaking it at home? A language that is no longer learned by children will lose its speakers as their generation ages. Teaching adults does not change that quickly. Reversing the trend means making new speakers, ideally children, faster than the old ones die. The toy model below uses two numbers: the share of speaker households that pass the language on, and the share of non-speakers who become speakers through schooling. It is invented for illustration and describes no real language. Interactive · Speakers over generations (toy model) Households that pass it on Non-speakers who learn it at school Each generation: speakers next = transmission × speakers now + school share × non-speakers now. Starts at 50% speakers; a generation is about 25 years. Invented parameters; real languages also lose speakers to migration, stigma and policy. Without JavaScript: with 70% transmission and no schooling, the share of speakers falls from 50% to about 3% in 8 generations; adding 20% school learning holds it at about 40%. Four revivals, four different roads Revitalisation has worked in places, and the successes do not look alike. Choose one below. Every figure is dated, and the figures count different things (census self-reports, for instance, are not tests of fluency). Interactive · Four cases Without JavaScript: Hebrew was revived as an everyday spoken language from the 1880s; Māori language nests (kōhanga reo) began in 1982; Welsh had 538,298 census speakers in 2021 (17.8%); Peru recognised Quechua as an official language in 1975. Put side by side, they show that no single lever works. Hebrew had a state to make it the language of school and government; Māori began with families and community-run language nests, and law followed; Welsh had a large base and consistent policy, yet its census share has not simply risen; Quechua has millions of speakers, so the problem is prestige and public use rather than numbers. Revival also changes the language: Israeli Hebrew shows influence from Yiddish and other languages its first speakers knew, which is why some linguists call it a new, mixed language rather than a resurrection, and others insist it is the same language continued. Both views are held by serious scholars. Writing systems: sound, syllable, word Writing is much younger than speech and most languages have never been written. It was invented independently at least three times, probably four: in Mesopotamia about 3400 to 3100 BCE, in Egypt about 3250 BCE (whether Egypt owed anything to Mesopotamia is debated, and Wikipedia's survey says no definitive determination has been made), in China by about 1250 BCE (the oracle bones), and in Mesoamerica, where Maya inscriptions begin by the 3rd century BCE. The systems differ in what one symbol stands for. TypeA symbol stands forExamples Alphabeta consonant or a vowelLatin, Greek, Cyrillic Abjada consonant (vowels optional or implied)Arabic, Hebrew Abugidaa consonant plus a default vowel, modified for othersDevanagari (Hindi), Ge'ez (Amharic), Canadian Aboriginal syllabics (Cree, Inuktitut) Syllabarya whole syllableJapanese kana, Cherokee Logographica word or morphemeChinese characters (with a sound clue inside most), Maya glyphs (a mix) Featurala feature of the soundHangul (Korean) Interactive · The same sound in eight scripts If a symbol shows as an empty box, your device lacks a font for that script; the text description still applies. The vowel marks are shown with their consonant. The most unusual is Hangul. It was created at the order of King Sejong of Joseon, presented to the court in 1443 and promulgated in 1446 in the Hunminjeongeum, with a commentary that explains the design: five basic consonant letters were modelled on the speech organs (ㄱ on the tongue root, ㅁ on the mouth, ㅇ on the throat) and the rest were derived by adding strokes. Letters are stacked into blocks, each with a starting consonant, a vowel and an optional final consonant. With 19 initials, 21 vowels and 28 final options (including none), the system can in principle form 19 × 21 × 28 = 11,172 blocks, of which about 2,000 are in common use. Computers use that same arithmetic. Try it. Interactive · Build a Hangul syllable Starting consonant Vowel Final consonant Block = 0xAC00 + (initial × 21 + vowel) × 28 + final, the formula Unicode itself uses. Romanisation follows the Revised Romanization of Korean in simplified form. Without JavaScript: ㅎ + ㅏ + ㄴ makes 한 (han); ㄱ + ㅡ + ㄹ makes 글 (geul); together they spell 한글, Hangul. What a graduate just sees There is no single count of languages: it depends on where you draw the line between a language and a dialect, and politics helps draw it. Families are shown by sound correspondences, not by resemblance, and some languages (isolates) have no known relatives. The biggest families by speakers (Indo-European) are not the biggest by languages (Atlantic–Congo, Austronesian). A typological "universal" is a strong tendency counted across languages, with exceptions that are real. Features go together (verb and object order predict adposition order) but languages are free to break the pattern. A language dies when children stop learning it; reviving it means making new speakers, and the successful cases take different roads. Scripts differ in what a symbol stands for, and none is more advanced than another. Where it shows up next Historical Linguistics (36/25): how families are proved, dated and drawn, and what the homeland debates rest on. Sociolinguistics (36/26): language policy, official status, who gets schooling in which language, and why revival campaigns succeed or stall. Syntax (36/23): explanations for the word-order harmony counted here. Computational Linguistics (36/30): most of the 7,000 languages have little text online, so language models are strongest in a few dozen languages. In the world: census questions on language, the design of keyboards and fonts, field recordings with community archives, and the UN's International Decade of Indigenous Languages (2022 to 2032). If you want to go deeper 📘 Textbook. Bernhard Comrie, Language Universals and Linguistic Typology (2nd edition, University of Chicago Press, 1989), the classic; or Viveka Velupillai, An Introduction to Linguistic Typology (John Benjamins, 2012), a more recent survey. 📘 Gentler. Nicholas Evans, Dying Words: Endangered Languages and What They Have to Tell Us (Wiley-Blackwell, 2010), written for general readers. 📘 Writing. Peter Daniels and William Bright (eds.), The World's Writing Systems (Oxford University Press, 1996), the standard reference. 🌐 Data you can explore. WALS Online and Glottolog, both free. 🌐 Free text. Essentials of Linguistics, 2nd edition (McMaster University), free online: ecampusontario.pressbooks.pub/essentialsoflinguistics2. 🌐 Free course. MIT OpenCourseWare 24.900 Introduction to Linguistics (Spring 2022), whose topics include language change and endangerment. The takeaway The world has about seven thousand languages, and how many exactly depends on a line that linguists and politics draw together. They fall into families, some of them lopsidedly large, and they differ in measurable ways, word order among them, that tend to go together without being laws. Many are being lost, because the one thing that keeps a language alive is children learning it, and that is also what revival has to rebuild. Try this: find out which languages are spoken around your own street or school, and which of them are written, taught in school and used on signs. The gaps between the three are your local picture of language vitality. Sources & Further Reading Sources. Language counts: Ethnologue 29th edition (2026) 7,170 and 27th edition (2024) 7,164, as reported in Wikipedia's "Ethnologue" article; Glottolog version 5.3, 2 March 2026, 7,674 spoken L1 languages, 246 families, 183 isolates (glottolog.org). Family sizes: Wikipedia's "List of language families" (languages from Glottolog 5.0, speakers from Ethnologue 2019). Weinreich 1945: the remark was reported to him by a listener; see YIVO. Heggarty et al., Science 381 (2023), as summarised in the University of Bamberg release and a critical review by Kassian and Starostin (2023). WALS Online: chapter 81 (Dryer; 564, 488, 95, 25, 11, 4, 189; total 1,376) and chapter 95 (Dryer; 472, 14, 42, 456; 158 others; total 1,142). Greenberg, "Some universals of grammar with particular reference to the order of meaningful elements" (1963). Hixkaryana toto yonoye kamara: Wikipedia, "Hixkaryana language" (600 native speakers in 2012). Malagasy Namonjy azy aho: Wikipedia, "Verb–object–subject word order". Endangered languages: UNESCO Atlas of the World's Languages in Danger (2010; 2,473 languages); Fishman, Reversing Language Shift (1991); Wikipedia's "Endangered language" for the 50 to 90% projection range. Hebrew: Wikipedia's "Revival of the Hebrew language" (1881, 1882, 1916–18 census 40%, 1922, 1948 80.9%, 2018 law); Zuckermann on Israeli Hebrew. Māori: Wikipedia's "Māori language" (kōhanga reo 1982, Māori Language Act 1987, 2018 census 185,955 or 3.96% of 4,699,755; its 2023 census figures conflicted on that page, so the 2018 count is used). Welsh: Wikipedia's "Welsh language" (2021 census 538,298 or 17.8%; 2011 562,016 or 19.0%; Welsh Language (Wales) Measure 2011; target of one million speakers by 2050; Welsh Language Act 1993). Quechua: Wikipedia's "Quechuan languages" (Adelaar 2004 estimate of 8 to 10 million; Peru 2017 census 13.9%; Peru official 1975, Bolivia 2009). Writing: Wikipedia's "History of writing"; Hangul: Wikipedia's "Hangul" (1443, 1446, 11,172 and about 2,000); Cherokee: Wikipedia's "Cherokee syllabary" (Sequoyah, 85 characters, Ꭶ = ga, Ꭷ = ka). UN International Decade of Indigenous Languages 2022–2032 (UNESCO). The toy models and the numbers on this page were checked in node. This piece describes languages and counts and is not a policy recommendation. 📘 Evans, Dying Words 📘 Daniels and Bright, The World's Writing Systems 🌐 WALS Online Step 4. Syntax: Sentences Are Trees, Not Strings ------------------------------------------------ Why: How words combine into sentences. You hear a sentence one word at a time, but you understand it as nested groups. Syntax is the study of those groups, the rules that build them, and the long argument over where the rules come from. Take a sentence that newspapers have run for decades: Police help dog bite victim. Read it fast and you picture a dog biting a person, and police helping the person. Read it again and you can see a different one: police helping a dog, a dog that is also a bite victim. The words did not change. What changed is which words you bundled together. That is the first discovery of syntax. A sentence is not a line of beads. It is a set of groups inside groups, and the meaning depends on the grouping. Every speaker builds those groups without effort, in every language ever studied, and almost nobody can say how. A person who has studied syntax keeps one habit for life: when a sentence is odd, ambiguous or hard, look for the hidden grouping. Syntax is also a battlefield. Is the machinery that builds the groups inborn and specific to language, or learned from use like any skill? The last sections give both sides in their own best terms. The course in one box Who takes itLinguistics majors, usually in year 2, and computer science, philosophy and language-teaching students as an elective. The Linguistic Society of America's guidance on the major, the UK QAA benchmark statement for linguistics and MIT's Course 24 all list syntax as core. Most courses teach the generative tradition (Chomsky and after), then compare other frameworks. What it is forDescribing how words combine into phrases and sentences, and testing a theory of that against what speakers accept and reject. Comes afterMorphology (36/22), the study of how words themselves are built. Leads toSemantics & Pragmatics (36/24), Language Acquisition (36/27), Psycholinguistics & the Brain (36/28) and Computational Linguistics (36/30). What this piece gives youYou will not be able to draw a full theory-specific tree. You will see why trees, why they loop, why some questions are unspeakable, why languages mirror each other, and what the fight over universal grammar is about. Easier first reads: Grammar Without Tears (26/08) and How Children Learn Language (26/06). Constituency: words group before they combine Linguists call a group of words that behaves as a unit a constituent. The test for one is practical. A constituent can usually be replaced by a single word of the right kind, moved as a block, or given as the answer to a question. In The cat sat on the mat, the mat can become it, and on the mat can become there. But sat on cannot be replaced by anything, and cat sat cannot be an answer to any question. Those two strings are not constituents, even though they are neighbours. a phrase = a head + whatever it takes along with itthe head decides the kind: a noun heads a noun phrase (NP), a verb a verb phrase (VP), a preposition a prepositional phrase (PP) Groups sit inside groups, which is what makes a tree. The same seven words, I saw the man with the telescope, give two trees. If with the telescope belongs to the verb phrase, you used the telescope to see. If it belongs to the man, the man had the telescope. Ambiguity of this kind is the signature of constituency: two groupings, one string. Interactive · Syntax tree builder Constituency test: tap a first and a last word. Is that stretch a constituent in this tree? Tap two words (or one twice). Trees are simplified (flat noun phrases, no X-bar layers); real theories draw more levels. Labels: S sentence, NP noun phrase, VP verb phrase, PP prepositional phrase, AP adjective phrase. Without JavaScript: in I saw [the man [with the telescope]] the man has the telescope; in I [saw the man] [with the telescope] the telescope is the instrument. How many ways could seven words be grouped in pairs? The count is the Catalan number: 132 binary shapes for seven words. The grammar of English allows two of them for this sentence. Nearly everything a syntax course teaches is a reason why the other 130 are not on offer. Recursion: a finite rule book, endless sentences A noun phrase can contain a noun phrase. The dog that chased the cat that ate the rat that lived in the house is one noun phrase with three more noun phrases tucked inside. Nothing stops you from adding a fourth. That is recursion: a rule that can apply to its own output. It is why a finite list of rules, and a finite vocabulary, still produce a language with no longest sentence. If someone gives you what they think is the longest sentence, you can add I think that at the front. Recursion comes in two shapes, and they behave very differently in a human head. Right-branching nests at the end: the next clause hangs off the last noun. Centre-embedding nests in the middle: the subject of the outer clause must wait, still unfinished, while a whole sentence is processed inside it. The grammar permits both equally. People manage the first for as long as you like and fail at the second after about two layers. George Miller and Noam Chomsky made this point in 1963: what the grammar allows and what a mind can handle are different questions. Linguists call them competence and performance. Interactive · Pile on the clauses Right-branchingCentre-embedded Clauses nested inside the first one Invented sentences. The "subjects waiting" count is how many subject phrases have been heard but have no verb yet at the moment the first verb arrives; it is the load your memory must carry. Without JavaScript: right-branching, The dog barked at the cat that chased the rat, is easy at any length. Centre-embedded, The dog the cat the rat bit chased barked, is grammatical and almost unreadable. Hauser, Chomsky and Fitch (2002) proposed that recursion, joining pieces into larger pieces, may be the one part of the language faculty that is unique to humans. That claim sparked the Pirahã argument below. Movement and dependency: the word that left a gap In What did Maria buy? the word what stands at the front, but it is understood as the thing bought. Compare Maria bought a book. The object is missing from its usual place. Syntacticians describe this as a filler (what) tied to a gap (after buy). Early generative grammar said the word moved. Other frameworks (HPSG, LFG, many construction-based accounts) say nothing moves; the two positions are simply linked. All of them must explain the same facts. The filler and the gap can be far apart: What do you think that Sam said that Maria bought? has five words between them. But not anywhere. In 1967 John "Haj" Ross showed that certain structures are islands: a gap inside them cannot be linked to a filler outside. A question that would need to reach into a relative clause, or into a clause that already has its own whether, sounds wrong, even though you understand what it asks. The verdicts below are the standard textbook ones. They are judgments, and experiments confirm the effects while finding them graded; some researchers (Hofmeister and Sag 2010 among them) argue that processing difficulty, not a rule, explains part of the pattern. Interactive · Predict, then reveal Does each question sound fine to a typical English speaker? The mark ___ shows where the answer would go. Judge the five questions. Without JavaScript: What did Maria buy ___? and What do you think Maria bought ___? are fine. What did you wonder whether Maria bought ___?, Who did Maria meet the man who kissed ___? and Whose did you read ___ book? are not. Languages differ in what counts as an island. Russian allows an equivalent of the last example, which English forbids. In Mandarin, Japanese and Korean the question word usually stays where an ordinary noun would be: Ni mai-le shenme? (2SG buy-PFV what, "What did you buy?"). In those languages the dependency is still there, but it is hidden in the meaning and not shown in the word order. Head direction: the same shape, mirrored Every phrase has a head, and the head sits on one side of what it takes along. In English the verb comes before its object (read the book), the preposition before its noun (in the house) and the noun before its relative clause (the book that I read). Japanese does the reverse on all three. This is a head-initial language mirrored by a head-final one. Taro-ga hon-o yon-daTaro-NOM book-ACC read-PST"Taro read the book." (Japanese: object before verb) Gwelodd Emrys y cisaw.3SG.PST Emrys the dog"Emrys saw the dog." (Welsh: verb first) Ali kitab-ı oku-duAli book-ACC read-PST"Ali read the book." (Turkish: object before verb) Joseph Greenberg noticed in 1963 that the orders travel together: languages that put the object before the verb overwhelmingly use postpositions (a small word after its noun, like Japanese Tōkyō kara, "Tokyo from"), and verb-first languages overwhelmingly use prepositions. The World Atlas of Language Structures (WALS) counts 1,376 languages for basic order, and the two big types lead it: subject–object–verb (564) and subject–verb–object (488), far ahead of verb-first orders (120). Interactive · Mirror, mirror Pick a language Language rows are simplified descriptions of the dominant pattern. Bars: WALS feature 81A, Dryer's 2013 chapter, as retrieved in 2026. Without JavaScript: English, Welsh and Swahili put verbs, prepositions and relative clauses after their heads; Japanese, Turkish, Korean and Quechua put them before. WALS counts 564 SOV, 488 SVO, 95 VSO languages. It is a tendency, not a law, and Mandarin and German are the standard cautions. Mandarin puts the verb before its object but its relative clause before the noun; German puts the verb last in subordinate clauses and second in main clauses. Chomsky's Principles and Parameters (1981) pictured head direction as one switch, set by a child from the first sentences it hears. Critics note that real languages mix more than a single switch can allow, and that typologists find statistical leanings, not on-off rules. Michael Dryer's counts across thousands of languages are the standard evidence. Structure dependence: what no child gets wrong Here is the best-known argument in the field. To make a question from The man is happy, move is to the front. A child could guess the rule is linear: take the first is and move it. That works for simple sentences. Try it on The man who is tall is happy. The first is gives Is the man who tall is happy?, which no English speaker says. The right rule refers to structure: move the is of the main clause, the one that is not buried inside who is tall. Interactive · Two rules, one question Rule A, linear: front the first auxiliary. Rule B, structural: front the auxiliary of the main clause. The page applies both and labels the output; the grammaticality labels are the standard textbook ones. Without JavaScript: from The man who is tall is happy, the linear rule gives Is the man who tall is happy? (wrong) and the structural rule gives Is the man who is tall happy? (right). Chomsky argued (from 1971, and often since) that a child would almost never hear a sentence that tells the two rules apart, yet children never try the linear one. Crain and Nakayama (1987) tested children aged three to five and found they did not produce the linear errors. Hence the "poverty of the stimulus": the input underdetermines the rule, so something inborn must steer the learner to structure. The reply from usage-based researchers has three parts. Pullum and Scholz (2002) argued that children hear more relevant sentences than the argument assumed. Reali and Christiansen (2005) showed that a simple statistical model trained on child-directed speech could prefer the right questions without any explicit rule. Kam and colleagues (2008) replied that the model succeeds partly by accident of the sentences chosen, and Berwick, Pietroski, Yankama and Chomsky (2011) reviewed the whole line. The argument is not over. What both sides accept is the result: children's questions respect structure. Universal grammar or learning from use? The fair version The central dispute can be stated without caricature. The generative view, from Noam Chomsky's Syntactic Structures (1957) onward, holds that humans are born with a language-specific capacity, universal grammar. Its evidence: structure dependence, the speed and uniformity of acquisition, and cases such as Nicaraguan Sign Language, where deaf children who were brought together from 1977 built a new language of their own. In the later Minimalist programme (from 1995), the claim shrank to something small: one combining operation, Merge, plus interfaces to thought and sound. The usage-based view, associated with Michael Tomasello (Constructing a Language, 2003), Adele Goldberg and Joan Bybee, holds that children learn language the way they learn other things: by general abilities for pattern-finding and for reading other people's intentions, applied to a great deal of input. Language is a large inventory of constructions, from words to partly filled templates, and abstraction grows from use. Evans and Levinson (2009) added a typologist's challenge: the supposed universals have counter-examples among the world's languages, so the real lesson is diversity. GenerativeUsage-based Core claimLanguage has a specialised, inborn coreGeneral learning builds language from use Best evidenceStructure dependence; new sign languagesGradual, item-by-item acquisition; statistical learning Hard questionWhat exactly is innate, and is it testable?Why do children avoid logically possible errors? Daniel Everett's work on Pirahã, a language of a small group on the Maici River in Brazil, put the argument in the news. In 2005 Everett claimed that Pirahã has no embedded clauses, no recursion in the grammar, and no numerals or colour words, and that its culture's focus on immediate experience explains this. If true, he said, recursion cannot be a universal property of grammar. Andrew Nevins, David Pesetsky and Cilene Rodrigues replied in 2009 that his data could be analysed otherwise, and that even if a language never uses recursion, an inborn capacity for it would not be refuted. Everett answered in turn. A 2016 corpus study by Futrell and colleagues found no clear evidence of embedding in the texts available. Chomsky has said the claim, even if right, would not touch his theory. The fieldwork is hard to repeat, few outsiders speak the language, and the matter is still not settled. What large language models show here is itself disputed: they learn many structural rules from text alone, which some read as support for learning from use, and others note that they see far more text than any child. Computational Linguistics (36/30) picks this up. Most working linguists today sit between the poles, accepting some inborn bias while arguing about how language-specific it is. What a graduate just sees Sentences are nested groups; ambiguity is two groupings of one string. A constituent is something you can replace with one word, move as a block, or give as an answer. Recursion is why there is no longest sentence, and memory, not grammar, is why we avoid centre-embedding. A "moved" word leaves a gap, and some regions (islands) will not let a dependency out. Head-initial and head-final languages are mirror images, and the mirror is a strong tendency, not a law. Children's grammar is built on structure, not on word order alone. "Universal grammar" is a claim to be tested, and each side has a real answer and a real hard case. Where it shows up next Semantics & Pragmatics (36/24): meaning is computed along the tree, so the grouping in I saw the man with the telescope decides the truth conditions. Language Acquisition (36/27): the poverty-of-the-stimulus argument and its rivals are tested on children's speech. Psycholinguistics & the Brain (36/28): centre-embedding and garden-path sentences show how real-time parsing works. Computational Linguistics (36/30): parsers, treebanks and the question of whether a language model has anything like a tree inside it. In the world: spell-checkers and grammar tools, translation between head-initial and head-final languages, legal drafting where a misplaced modifier costs money. If you want to go deeper 📘 Textbook. Andrew Carnie, Syntax: A Generative Introduction (Wiley-Blackwell, 4th edition 2021), with worked exercises; or David Adger, Core Syntax: A Minimalist Approach (Oxford, 2003), a tighter, more theoretical route. Check for a current edition. 📘 Gentler. Steven Pinker, The Language Instinct (1994), a lively case for the generative side; then Daniel Everett, Don't Sleep, There Are Snakes (2008), and Michael Tomasello, Constructing a Language (2003), for the other voices. 🌐 Free text. Catherine Anderson and others, Essentials of Linguistics, a free open textbook with a syntax chapter, and the free World Atlas of Language Structures at wals.info. 🌐 Free course. MIT OpenCourseWare 24.900 Introduction to Linguistics (Spring 2022), and for a full syntax course 24.902 Language and Its Structure II: Syntax (Fall 2003). The takeaway Sentences are built from nested groups, and the grouping carries the meaning. The same few devices, heads, recursion and gaps, appear in every language, arranged in mirror-image ways. Where they come from, a language-specific inborn core or general learning applied to experience, is still argued by serious people, with data on both sides. Try this: find a headline or sign with two readings (try "Kids make nutritious snacks") and write down the two groupings. Then try the test: which words can be replaced by one? Sources & Further Reading Sources. Chomsky, Syntactic Structures (1957), Lectures on Government and Binding (1981), The Minimalist Program (1995); Ross, Constraints on Variables in Syntax (1967); Miller and Chomsky 1963; Greenberg, "Some universals of grammar with particular reference to the order of meaningful elements" (1963); Hauser, Chomsky and Fitch, "The faculty of language," Science 2002; Crain and Nakayama, Language 1987; Pullum and Scholz, The Linguistic Review 2002; Reali and Christiansen, Cognitive Science 2005; Kam et al., Cognitive Science 2008; Berwick, Pietroski, Yankama and Chomsky, Cognitive Science 2011; Everett, "Cultural constraints on grammar and cognition in Pirahã," Current Anthropology 2005; Nevins, Pesetsky and Rodrigues, "Pirahã exceptionality: a reassessment," Language 2009; Futrell et al., PLoS ONE 2016; Tomasello 2003; Evans and Levinson, "The myth of language universals," Behavioral and Brain Sciences 2009; Hofmeister and Sag, Language 2010; WALS feature 81A (Dryer), word counts retrieved 2026. The Catalan-number count and all outputs of the interactives were computed in node. Example sentences are standard textbook or invented; Welsh, Japanese, Turkish and Mandarin examples were checked against grammars. This piece describes research and takes no side in the debate it reports. 📘 Carnie, Syntax: A Generative Introduction 📘 Tomasello, Constructing a Language 🌐 World Atlas of Language Structures Step 5. Semantics & Pragmatics: What Is Said and What Is Meant -------------------------------------------------------------- Why: What sentences mean, and what speakers mean by them. A sentence has a meaning that comes from its words. A speaker adds a meaning that comes from the occasion. Learning to pull those two apart is most of the subject. You ask a colleague whether the new hire is any good, and she answers: "Well, she's always on time." Nothing in that sentence says no. You understood no, instantly, and so did she, and if you challenged her she could say, truthfully, that she only mentioned punctuality. Two systems are at work. One computes what the sentence means whoever says it, when, and to whom: that is semantics. The other works out what this speaker, here, is getting at by saying it: that is pragmatics. A graduate of this course sees both layers under every conversation, and can tell a literal falsehood from a lie, a hint from a promise, a rude request from a polite one in another language. The course in one box Who takes itLinguistics majors, usually in year 2 or 3, and philosophy, computer science and psychology students as an elective. The LSA and QAA guidance and MIT's Course 24 all put meaning beside sound, words and syntax as a core area. Programmes differ on how much formal logic they ask for: some teach semantics with set theory and lambda notation, others stay with concepts and cases. What it is forSaying exactly what a sentence commits a speaker to, and explaining how listeners get far more than that. Comes afterSyntax (36/23), because meaning is built along the tree, and Philosophy of Language (36/07). Leads toLanguage Acquisition (36/27), Psycholinguistics & the Brain (36/28) and Computational Linguistics (36/30). What this piece gives youYou will not be able to write a formal semantic derivation. You will know six ideas that let you read ordinary speech more sharply. Easier first reads: Grammar Without Tears (26/08) and The Biases That Fool Everyone (17/02). Compositionality: meaning is built from parts You have never heard The purple elephant is next to the small square, and you understand it. That is only possible if the meaning of a whole sentence is computed from the meanings of its words and the way they are put together. This is the principle of compositionality, usually credited to Gottlob Frege, although scholars point out that he never stated it in this form. Richard Montague turned it into a working programme around 1970 by giving English sentences precise truth conditions, built step by step along their syntax. The picture to carry is that words denote sets, and grammar combines the sets. Circle picks out the circles in a scene. Red picks out the red things. Red circle is the overlap. Every is a relation between two sets: every circle is red is true if the circles are inside the red things; some asks only that the sets overlap; no asks that they do not overlap; most asks that the overlap is bigger than the rest of the first set. Barwise and Cooper (1981) showed that this view of quantifiers is not a quirk of English, and that most cannot be expressed in ordinary first-order logic at all. Interactive · A tiny world, a sentence, a truth value Tap a shape to change it: circle or square, then red or blue. Sentence Four objects, numbered 1 to 4 from the left. "Next to" means touching neighbours in the row. A toy, not a full grammar. Without JavaScript: in a row of red circle, red square, blue circle, blue square, Some square is blue is true, Every circle is red is false (circle 3 is blue) and Most shapes are red is false (two of four is not most). Compositionality has limits, and the interesting cases are the ones where it seems to break. In an idiom such as kick the bucket, the whole is stored as one unit. A red pen may write in blue, and red hair is not red like a fire engine, so adjectives shift with the noun they sit beside. Linguists keep the principle as the default and ask, case by case, which extra mechanisms (stored chunks, coercion, context) are needed. Entailment: what a sentence commits you to Wittgenstein wrote in the Tractatus (4.024): "To understand a proposition means to know what is the case, if it is true." That is the heart of truth-conditional semantics. If you know the conditions under which a sentence is true, you can say what follows from it. A sentence A entails B if B is true in every situation where A is true. All three students passed entails Some student passed, but not the reverse. A entails B when every situation that makes A true also makes B trueso the situations for A are a subset of the situations for B The tool below makes the situations literal. Three students either pass or fail, which gives 2×2×2 = 8 possible situations. Pick two sentences and see which situations make each true. Interactive · The eight situations Sentence A Sentence B A true   B true   both   P = passed, F = failed (students 1, 2, 3) Without JavaScript: All passed is true in 1 of 8 situations and Some passed in 7, and the one is among the seven, so All entails Some. Some does not entail All. Notice that Some passed often suggests that not all did. That is not an entailment, because a speaker can say "Some passed, in fact all of them did" without contradiction. It is an implicature, the subject of the next-but-one section. Presupposition: what a sentence takes for granted Mia stopped smoking says that she does not smoke now. It also takes for granted that she used to. The second is a presupposition, and it behaves differently from an entailment. Negate the sentence, Mia didn't stop smoking, turn it into a question, Did Mia stop smoking?, or tuck it under if, and the "she used to smoke" part stays. The entailment about now does not. This family-of-sentences test is how semanticists separate the two. The words that trigger presuppositions include stop, regret, the, possessives (Jo's sister) and clefts (It was Lena who broke the cup). Interactive · Does it survive? Sentence Without JavaScript: the presupposition (Mia used to smoke) survives not, a question and if. The entailment (Mia does not smoke now) does not. The oldest puzzle is The king of France is bald, said when France has no king. Bertrand Russell (1905) said the sentence is simply false, because it asserts that exactly one king exists and is bald. Peter Strawson (1950) said it is neither true nor false, because the presupposition fails and the question of baldness does not arise. Semanticists still divide on this. What everyone agrees on is the practical side. Listeners usually do not object to a presupposition; they quietly accommodate it. If a stranger says "Sorry I'm late, my sister's car broke down," you add that she has a sister. This is why presuppositions carry persuasion: "When did you stop cheating?" smuggles in the cheating. Projection is not absolute. If the presupposition has already been stated in the sentence, it disappears: If Mia used to smoke, she has stopped no longer commits the speaker to her smoking. The tool above shows this as the last option. Implicature: Grice and the cooperative speaker H. Paul Grice's 1967 William James lectures (published as "Logic and Conversation," 1975) started modern pragmatics. His idea: conversation is cooperative, and speakers and hearers both assume it. This Cooperative Principle unfolds into four groups of expectations: MaximExpectation QuantitySay as much as needed, and no more QualitySay only what you believe to be true RelationBe relevant MannerBe clear: avoid obscurity and needless wordiness The clever part is what happens when a speaker visibly breaks a maxim. The hearer assumes the speaker is still cooperating and looks for a reason. "She's always on time," in answer to "Is she any good?", is not relevant, unless the speaker has nothing better to say. That is the implicature. Grice listed tests: an implicature can be calculated from the maxims; it can be cancelled ("...but she's also brilliant"); and it hangs on what is said, not on the wording (a paraphrase carries it too). Interactive · Which maxim is being played with? Pick a maxim for each exchange. Invented exchanges. Some can be read two ways; the answer given is the usual textbook one. Without JavaScript: answering "How was the lecture?" with "The room was warm" flouts Relation; "Ms Ruiz dresses well and arrives on time", as a reference letter, flouts Quantity; "What lovely weather!" in a storm flouts Quality; a long, formal description of someone's singing flouts Manner. Not everyone holds that four maxims are the right foundation. Dan Sperber and Deirdre Wilson's Relevance Theory (1986) reduces the machinery to one principle: hearers look for the interpretation that gives the most effect for the least effort. Stephen Levinson and Laurence Horn each proposed a different, smaller set of principles. Experimental pragmatics, from Ira Noveck's 2001 work on children and some to many later studies, tests the predictions: adults often add the "not all" meaning, children more rarely, and it takes extra time. A large rival question, still open, is how much of the "not all" is a quick default and how much is built on the spot. Speech acts: saying is doing John Austin gave lectures at Harvard in 1955, published after his death as How to Do Things with Words (1962). His starting point was that many sentences are not reports but deeds. "I promise to be there," "I name this ship," "I bet you five euros." To say them, in the right circumstances, is to do the thing. He split every utterance into the words and grammar (locution), what the speaker does with them, such as warn, request or promise (illocution), and what happens to the hearer (perlocution). Deeds can fail in ways truths cannot. If a passer-by says "I now pronounce you married," nothing happens, because the speaker is not authorised. Austin called these failures infelicities; the conditions that make a deed work are its felicity conditions. John Searle (1969, 1976) sorted the force of utterances into five families: telling how things are (assertives), trying to get the hearer to do something (directives), committing the speaker (commissives), expressing a feeling (expressives) and changing the world by saying so (declarations). Interactive · What is being done? Pick a force for each utterance. Who says "I now pronounce you married"?A registrar or officiant in a ceremonyA guest, in the middle of the party Without JavaScript: "Can you pass the salt?" is grammatically a question and in force a directive (an indirect speech act). "I promise" is a commissive; "Congratulations!" an expressive; "You're fired" said by a manager, a declaration. Indirect speech acts show why pragmatics matters. Can you pass the salt? is nearly never about ability. Answering "Yes" and doing nothing is a joke because the hearer understood perfectly well and chose to be literal. Studies of requests across languages (the Cross-Cultural Speech Act Realization Project of Blum-Kulka, House and Kasper, 1989) found that every language they examined has direct and indirect ways to ask, but the mix differs. Politeness across cultures: the price of a request Why does the same request come out in so many shapes? Penelope Brown and Stephen Levinson (1978; book 1987) built on Erving Goffman's idea of face, a person's public self-image. Most acts, like requests or criticism, threaten someone's face. Speakers weigh the threat by three things: the social distance between them, the power of the hearer over the speaker, and how large an imposition the act is. Higher weight, more care. Their examples came from English, Tamil and Tzeltal, a Mayan language of Mexico. They claimed that the logic is universal, though what counts as a heavy imposition is not. weight of the threat = social distance + power of the hearer + size of the impositionBrown and Levinson's formula; the page's thresholds below are illustrative, not theirs Interactive · How hard is this request? Social distance (1 close, 7 strangers) Hearer's power over you (1 low, 7 high) Size of the favour (1 tiny, 7 huge) Language Request: "pass me the salt". Strategy bands are my own illustrative cut-offs along Brown and Levinson's scale of bald on record, positive politeness, negative politeness, off record. Translations of the sample requests are given in brackets. Without JavaScript: at weights of 3 to 7 the request is direct (English "Pass the salt"); at 12 to 15 it becomes elaborate ("Sorry to bother you, could you possibly pass the salt?"); from 16 up it is hinted, and at 20 or 21 it may not be made. The critics have a strong point. Yoshiko Matsumoto (1988) and Sachiko Ide (1989) argued that in Japanese, honorifics are largely required by the social relation (wakimae, "discernment") and not chosen to save face, so the model cannot explain them. Gu Yueguo (1990) made a similar argument about Chinese, and Anna Wierzbicka argued that each culture has its own "cultural scripts" (for example, Anglo reluctance to give flat orders compared with the more direct styles of Polish or Russian). Many languages also build status into grammar: the T–V choice of French tu/vous, Spanish tú/usted and German du/Sie (Brown and Gilman 1960), Korean speech levels, and the Javanese levels ngoko and krama. A fair summary: face is a useful and widely cross-checked idea, and a universal formula is not. What a graduate just sees The meaning of a sentence is computed from its parts and its tree; the meaning of an utterance is that, plus the occasion. "Every", "some", "most" and "no" are relations between sets, and you can check them against a picture. An entailment holds whatever the context; an implicature can be cancelled without contradiction. A presupposition survives negation and questions; if a sentence assumes something, argue with the assumption, not the sentence. Flouting a maxim is a signal, and the listener's job is to ask why. Saying can be doing, and a deed can fail because the wrong person said it. Politeness follows the weight of a request, and what counts as heavy differs by culture. Where it shows up next Language Acquisition (36/27): children treat some more literally than adults do, and slowly learn who may say what to whom. Psycholinguistics & the Brain (36/28): how fast listeners add implicatures, and what happens to them in some brain injuries. Computational Linguistics (36/30): textual entailment tasks, question answering and why chatbots must learn presupposition and indirect requests. Philosophy of Language (36/07) and Logic & Argument (36/01): truth conditions, the Russell-Strawson dispute and the logic of "if". In the world: contracts and warnings (what counts as "said"), advertising ("up to 50% off"), customer service scripts, and diplomacy. If you want to go deeper 📘 Textbook. Paul Kroeger, Analyzing Meaning (Language Science Press, free), and Betty Birner, Introduction to Pragmatics (Wiley-Blackwell, 2013). For formal semantics, Irene Heim and Angelika Kratzer, Semantics in Generative Grammar (Blackwell, 1998). Check for newer editions. 📘 Gentler. Steven Pinker, The Stuff of Thought (2007), which has chapters on indirect speech and politeness; and H. P. Grice, "Logic and Conversation," the original and short. 🌐 Free text. The open textbook Essentials of Linguistics has chapters on semantics and pragmatics. 🌐 Free course. MIT OpenCourseWare 24.900 Introduction to Linguistics (Spring 2022) has a semantics block (lectures 17 to 21). The takeaway Every utterance has a layer of meaning built from its words and tree, and a layer built from why it was said. Entailments are fixed by the sentence; presuppositions are what it takes for granted; implicatures are what a cooperative listener infers when a maxim seems broken. The rest, acts and politeness, is what we do with all three. Try this: next time someone answers a question sideways, write down what they literally said, what they implied, and which maxim they bent. Then ask whether they could deny the implication. Sources & Further Reading Sources. Frege and compositionality (see Janssen 1997 on the attribution); Montague, "English as a formal language" (1970); Barwise and Cooper, Linguistics and Philosophy 1981; Wittgenstein, Tractatus 4.024 (Ogden's 1922 translation); Russell, "On denoting" (1905); Strawson, "On referring" (1950); Grice, "Logic and Conversation" (1975, William James lectures 1967); Sperber and Wilson, Relevance (1986); Horn 1972 on scales; Levinson, Presumptive Meanings (2000); Noveck, Cognition 2001; Austin, How to Do Things with Words (1962); Searle, Speech Acts (1969) and "A classification of illocutionary acts" (1976); Blum-Kulka, House and Kasper, Cross-Cultural Pragmatics (1989); Brown and Levinson, Politeness (1987; first 1978); Goffman 1967; Brown and Gilman 1960; Matsumoto, Journal of Pragmatics 1988; Ide, Multilingua 1989; Gu, Journal of Pragmatics 1990; Wierzbicka, Cross-Cultural Pragmatics (1991). The 8-situation counts and every output of the interactives were computed in node. Example dialogues are invented; non-English requests were checked against grammars. This piece describes research and takes no side where the field is divided. 📘 Birner, Introduction to Pragmatics 📘 Heim and Kratzer, Semantics in Generative Grammar 🌐 MIT OCW 24.900 Step 6. Philosophy of Language: What Words Mean and What They Do ---------------------------------------------------------------- Why: Meaning and reference, from the philosophy side. "The Morning Star is the Evening Star" told astronomers something. "The Morning Star is the Morning Star" told them nothing. Yet both name the same planet. A century of philosophy grew from that puzzle, and from the stranger fact that we mean far more than we say. Someone asks, "Can you pass the salt?" and you do not answer "Yes." Someone says, "Some of the students passed," and you hear that not all did, although nobody said so. A friend says, "Great weather," in a storm, and you hear the opposite. In each case the words on the page are not what is communicated, and in each case you got it right without thinking. Philosophy of language asks how that is possible. How can a sound or a mark be about something? Why can a true statement be informative or empty depending on which name you use? What are we doing when we promise, or name a ship, or say "I do"? The questions are old, but the modern subject begins in the 1890s with Gottlob Frege, and the tools it built (sense, reference, presupposition, rigid designation, implicature) are now used in linguistics, law, computer science and the study of language models. The course in one box Who takes itPhilosophy majors, usually in year 3, at programmes in the analytic tradition (MIT's philosophy department, for one, teaches it as 24.251). Linguistics majors take its formal cousin, semantics and pragmatics, and it sits on both routes. What it is forSeeing exactly what a sentence says, what its speaker meant, and what each of them has to do with the world. Comes afterLogic & Argument (36/01), for the quantifiers and the idea of validity. Leads toSemantics & Pragmatics (36/24), the linguist's version; Philosophy of Mind (36/05), where meaning and thought are one question; Contemporary Philosophy (36/12); and Computational Linguistics (36/30), where "meaning" meets statistics. What this piece gives youYou will not be able to defend a theory of reference in a seminar. You will know the six or seven moves that every such theory is built from, why each was made, and what its critics say. Easier first reads: How Children Learn Language (26/06) and How Language Changes (26/05). Sense and reference: Frege and the two stars The obvious theory says a name means the thing it names. "Venus" means Venus. Then "Hesperus" (the evening star) and "Phosphorus" (the morning star) mean the same thing, because both are Venus, and the sentence "Hesperus is Phosphorus" should say no more than "Venus is Venus." But the first was a discovery, and the second is a triviality. In "Über Sinn und Bedeutung" (1892) Frege put it plainly: statements of the form a = a and a = b differ in what they tell you. His answer is that an expression has two things attached. Its reference is the object it picks out. Its sense is the way it picks it out, the route by which a thinker gets to the object. "The bright point at dawn" and "the bright point at dusk" are two routes to one planet. Frege then did the same for whole sentences: the sense of a sentence is the thought it expresses, and its reference is its truth-value. That explains why swapping co-referring names keeps a sentence true (the reference is unchanged), yet can change what a thinker believes (the sense is changed). sense → referencesense is the route; reference is where the route ends. Two routes can end at one place. Interactive · Which identity teaches you something? Pick the left-hand and right-hand expression of "A is B". A B Reference: Venus is the planet seen both at dawn and dusk; the Moon is a separate body. The verdict follows Frege's test (cognitive value), not a full theory. Without JavaScript: "The Morning Star is the Evening Star" is true and informative (one reference, two senses); "The Morning Star is the Morning Star" is true and trivial; "The Morning Star is the Moon" is informative and false. Sense does extra work. Ancient sailors did not know that the two stars were one, so "the sailors believed the Morning Star rose at dawn" is true while "the sailors believed Venus rose at dawn and again at dusk" may not be. Frege said that inside a belief report an expression refers to its usual sense, not to its usual object. Critics found this heavy, and the next two sections are attempts to do without it. Descriptions: Russell and the king of France Bertrand Russell's "On Denoting" (1905) attacked a different puzzle. Take "The present King of France is bald." France has no king. If the phrase must refer to something for the sentence to mean anything, the sentence is meaningless. But it is not: it is plainly false, or so Russell thought. His fix was to say that "the F is G" is not a subject and predicate at all. It is a hidden claim in three parts: there is an F  ·  there is only one F  ·  it is G"The King of France is bald" is false because the first part fails. This also exposes an ambiguity. "The King of France is not bald" can deny the whole claim (true: there is no such king) or say of the king that he lacks hair-loss (false: there is no king). Logicians call this a difference of scope. Peter Strawson replied in "On Referring" (1950) that the sentence has a presupposition, that there is a king, and when that fails the sentence is neither true nor false. Keith Donnellan (1966) added that "the man drinking a martini" can pick out a particular man even if he is drinking water, which Russell's account does not allow. Play with the scenarios below. Interactive · Who is right about the king? Computed from Russell's three clauses for "the King is bald". Strawson's column gives "no truth-value" whenever there is not exactly one king. Without JavaScript: with no king, Russell makes "the King is bald" false and "it is not the case that the King is bald" true; Strawson makes both lack a truth-value. Names: Kripke and the rigid designator Russell extended his analysis to ordinary names: "Aristotle" is short for a description such as "the teacher of Alexander". Frege agreed on something like it. In three lectures at Princeton in January 1970, later the book Naming and Necessity (1980), Saul Kripke argued that this is wrong. His arguments, in order: Modal. Aristotle might never have taught Alexander (his father was a court physician; he might have become one). If "Aristotle" meant "the teacher of Alexander", "Aristotle taught Alexander" would be necessary. It is not. A name is a rigid designator: it picks the same person in every possible world, while a description picks whoever fits in each. Epistemic. Most of us know little that is uniquely true of the people we name, and we still refer to them. Semantic. Suppose "Gödel" is associated only with "proved incompleteness", and the proof was actually done by a man called Schmidt. A description theory says that you are talking about Schmidt. Almost no one thinks so (Kripke's case is hypothetical; Gödel's proof was his own). Interactive · Names and descriptions across possible worlds Three invented worlds, labelled as such: in w1 Aristotle became a physician and Isocrates tutored Alexander; in w2 Alexander had no tutor. Everything outside w0 is a thought experiment. Without JavaScript: "Aristotle" picks Aristotle in every world; "the tutor of Alexander" picks Aristotle in w0, Isocrates in w1 and nobody in w2. So "Aristotle tutored Alexander" is true in w0 only, while "the tutor of Alexander tutored Alexander" is true in all three. What fixes reference instead? Kripke's picture is causal-historical: a baptism ties a name to a person, and each speaker who hears it and means to use it as others do passes the link along. Predict first, then look. Interactive · Predict, then reveal Suppose the only thing you know about "Gödel" is that he proved the incompleteness theorem, and (contrary to fact) Schmidt had proved it. Who are you talking about when you say "Gödel was a genius"? Schmidt, the man who fits the descriptionGödel, the man at the other end of the chain Pick one, then read the answer. Without JavaScript: the description theory says Schmidt; Kripke says Gödel, because your use of the name is linked back to Gödel by a chain of speakers. The critics have good cases. Gareth Evans (1973) noted that "Madagascar" began as a name of part of the African mainland and came, through a chain of speakers' mistakes, to name the island, so the chain alone can't be all of it. And Kripke left an old puzzle open: if a name just stands for its bearer, why is "Hesperus is Phosphorus" informative? Defenders of the Millian view reply with pragmatics; Fregeans keep sense. The PhilPapers Survey of 2020 asked professionals exactly this ("Proper names: Millian or Fregean?"); see its results page for the split, which I do not reprint here. Meaning as use: Wittgenstein's two philosophies Ludwig Wittgenstein held both a view of language and its sharpest critique. The Tractatus Logico-Philosophicus (German 1921, English 1922) said that a sentence is a picture of a possible fact, names stand for simple objects, and what can't be pictured can't be said. "The limits of my language mean the limits of my world."Wittgenstein, Tractatus 5.6, C. K. Ogden's 1922 translation Later he rejected the picture. Philosophical Investigations (published 1953, after his death in 1951) starts with a scene: a builder calls "Slab!" and an assistant brings a slab. Here a word is not a label for a thing; it is a move in a language-game, tied to what people do. For a large class of cases, he wrote, the meaning of a word is its use in the language (section 43). He also argued that "game" has no single feature common to all games, only overlapping family resemblances (section 66), and that a language only one person could ever understand is incoherent (the private language argument, from about section 243). Each claim has opponents, and the last has produced shelves of books. What most readers keep is a change of question: not "what does this word stand for?" but "what are people doing with it?" Doing things with words: Austin and Searle J. L. Austin's William James Lectures at Harvard in 1955, published in 1962 as How to Do Things with Words, began with an observation. Some sentences are not true or false; saying them does something. "I name this ship," "I promise," "I bet you," "I do." Austin called these performatives, and noticed that they can go wrong in different ways from statements. They can be unhappy rather than false. Each needs a recognised procedure, the right people in the right circumstances, correct execution, and (for some) sincerity. The tool below lets you break them. Interactive · "I now pronounce you married" Speaker has the authority The couple are eligible and willing It is a stage play Speaker means it Austin's categories: misfire (a condition on the procedure fails, so the act is void) and abuse (the act is done but insincerely). Stage use he called "etiolated". Marriage law differs by country; this is Austin's idea, not a legal test. Without JavaScript: if the speaker has no authority the act misfires and no marriage results; if everything is in order but the speaker is insincere the marriage still happens (an abuse); on stage nothing is done. Austin's second step was to say that every utterance is several acts at once: the locutionary act (uttering words with a sense), the illocutionary act (what you do in saying it: warn, ask, order) and the perlocutionary act (what you bring about by it: scare, persuade). "There's a bull in that field" can be a warning, whether or not it frightens anyone. John Searle's Speech Acts (1969) developed rules for them, and his later scheme sorts illocutions into five kinds: assertives, directives, commissives, expressives and declarations. "Can you pass the salt?" is the case that shows why this matters: it is literally a question about ability, and in practice a request. What you mean beyond what you say: Grice Paul Grice, in the William James Lectures of 1967 (the essay "Logic and Conversation" appeared in 1975), explained the salt case with one idea. Conversation is cooperative, and we assume a speaker is following four maxims: say as much as is needed (Quantity), say what you believe true (Quality), be relevant (Relation), be clear and orderly (Manner). When a speaker visibly breaks one, we infer that something else is meant: an implicature. A reference letter that praises a student's handwriting and nothing else says very little; the hearer concludes there is nothing good to say. Sort the cases yourself. Interactive · Which maxim is being flouted? Pick a maxim. Next case Invented cases in the spirit of Grice's. Score is kept on this page only. Without JavaScript: the vague letter flouts Quantity; "Great weather!" in a storm flouts Quality; a reply that seems off-topic flouts Relation; an absurdly roundabout phrase flouts Manner. Implicatures differ from what a sentence strictly entails by a test. They can be cancelled: "Some of the students passed; in fact all did" is not a contradiction, so "not all" was never part of the meaning of "some", only something we read in. Rivals disagree on how many maxims there are. Dan Sperber and Deirdre Wilson's relevance theory (1986) reduces them to a single principle of relevance, and the neo-Griceans (Larry Horn, Stephen Levinson) keep a short list. India's argument about words: Nyāya, Mīmāṃsā and Buddhist apoha None of this was invented only in Europe. Indian philosophers argued about meaning for well over a thousand years, often more precisely than the West did for most of that time. The main schools were the Nyāya logicians (the Nyāya Sūtra, compiled in the early centuries CE), the Mīmāṃsā interpreters of the Vedic texts (Jaimini's Sūtra, with Śabara, and later Kumārila Bhaṭṭa and Prabhākara, about the 7th century CE), the Buddhist logicians Dignāga and Dharmakīrti, and the grammarians led by Bhartṛhari. Four debates are worth knowing. What does a word denote? The Nyāya answer, in the Sūtra: a word denotes an individual together with its form and its universal (a cow, the shape of a cow, cowness). Others stressed the universal alone; Buddhists denied universals exist. Where does word-meaning come from? Mīmāṃsā held that the link between word and meaning is eternal and not made by anyone; this protected the Veda as authorless. Nyāya held that it is conventional, set by divine will (the later Navya-Nyāya school, from Gaṅgeśa in the 14th century, kept and refined that view). How do we learn it? By watching elders. If one elder says "Bring the cow" and another does it, a child links the words to the act, then isolates "cow" by changing one word. Many scholars note the likeness to Wittgenstein's builders; it is an echo, not an influence. Do words mean on their own, or only in sentences? Kumārila: words first denote their meanings, then the sentence's meaning is put together from them. Prabhākara: words convey meanings only as already related, which he took to be what children learn from commands. Bhartṛhari went further: the sentence is a whole (sphoṭa), understood in a flash. The Buddhist apoha theory says a word like "cow" works by excluding what is not a cow, so it needs no real universals. Nyāya textbooks also list conditions for a string of words to give an understood sentence: ākāṅkṣā (expectancy: the words need each other to complete a sense), yogyatā (compatibility: the meanings can sensibly combine) and sannidhi (proximity: said together, not hours apart). This is a surprisingly modern checklist, and you can run it. Interactive · Does this string become a sentence? The "sprinkles with fire" example is the standard classical illustration of failed compatibility. The verdicts are the traditional ones, applied to simple examples. Without JavaScript: "Bring the cow" meets all three; a bare list of nouns lacks expectancy; "he sprinkles with fire" lacks compatibility; words said an hour apart lack proximity. What a graduate just sees Two expressions can refer to the same thing and still not mean the same, and the proof is that one identity is news and the other is not. A sentence that sounds like it is about someone may secretly be a claim that "there is exactly one such person", and where "not" goes can change the truth. A name is not a hidden description; it is a tag that stays on the person across the ways things could have gone. Ask what people are doing with a word before asking what it stands for. Some sentences are acts: they can misfire without being false. What is meant often outruns what is said, and the extra can be cancelled, which shows it was never part of the sentence. The theory of meaning has been argued across traditions, and the Indian debates are not footnotes. Where it shows up next Semantics & Pragmatics (36/24): the linguist's formal versions of sense, scope, presupposition and implicature. Philosophy of Mind (36/05): whether thoughts are meaningful in the way words are, and what a machine that handles words understands. Computational Linguistics (36/30) and How Large Language Models Actually Work (29/02): models learn meaning as use, in Wittgenstein's sense, and whether that is all of meaning is exactly the debate above. In the world: contracts and statutes (a promise is an act; "the" in a statute may bring a presupposition); advertising and politics (implicature lets a speaker mislead while saying only true things); and courtroom questions about what a defendant "meant". If you want to go deeper 📘 Textbook. William Lycan, Philosophy of Language: A Contemporary Introduction (Routledge, 3rd edition, 2019), clear and organised by problem; Alex Miller, Philosophy of Language (Routledge, also in later editions), similar in range. Check the current edition. 📘 Primary and gentler. Bertrand Russell's The Problems of Philosophy (1912), chapter 5, free at Project Gutenberg, for the idea of knowledge by description; then Saul Kripke, Naming and Necessity (Harvard University Press, 1980), which is short and readable. 📘 India. Bimal Krishna Matilal, The Word and the World: India's Contribution to the Study of Language (Oxford University Press, 1990). 🌐 Free course. MIT OpenCourseWare 24.251 Introduction to Philosophy of Language (Fall 2011), with lecture notes. The Stanford Encyclopedia of Philosophy has free entries on descriptions, implicature and speech acts. The takeaway Words have a route (sense) and a destination (reference); sentences can hide structure; names tag people rather than describe them; words are moves in activities; some utterances are acts; and speakers mean more than they say. Each of these was introduced to solve a particular puzzle, and each still has live critics. Indian philosophers worked on the same puzzles from different starting points. Try this: for one day, catch three things people communicate without saying them (a reply that skips the question, a "fine", a "some"), and name the maxim that makes you hear it. Sources & Further Reading Sources. Frege, "Über Sinn und Bedeutung" (1892); Russell, "On Denoting", Mind 1905; Strawson, "On Referring", Mind 1950; Donnellan, "Reference and Definite Descriptions", Philosophical Review 1966; Kripke, Naming and Necessity (lectures January 1970; book 1980); Evans, "The Causal Theory of Names", 1973; Wittgenstein, Tractatus (1921; Ogden 1922) and Philosophical Investigations (1953); Austin, How to Do Things with Words (lectures 1955; book 1962); Searle, Speech Acts (1969); Grice, "Logic and Conversation" (William James Lectures 1967; published 1975); Sperber and Wilson, Relevance (1986); Stanford Encyclopedia of Philosophy entries on these topics; Matilal (1990) for the Indian theories. The PhilPapers Survey 2020 (philpapers.org) asks about proper names. 📘 Lycan, Philosophy of Language 📘 Kripke, Naming and Necessity 🌐 MIT OCW 24.251 Step 7. Historical Linguistics: Sound Laws and the Family Trees They Draw ------------------------------------------------------------------------- Why: How languages change and how we reconstruct their ancestors. How a handful of regular sound changes let us read languages nobody has spoken for six thousand years, and what happens when the evidence runs out. Latin pater, Spanish padre, Sanskrit pitā, English father. Latin pēs, Greek pous, English foot. Latin piscis, English fish. Latin plēnus, English full. Where Latin and Greek have a p, English has an f, and it does so again and again. A coincidence would not repeat itself that neatly. It is not a coincidence, and the discovery that it is not is the founding moment of historical linguistics. In 1786 Sir William Jones, a judge in Calcutta, told the Asiatic Society that Sanskrit, Greek and Latin were so alike in roots and grammar that they must have "sprung from some common source". But resemblance was the easy part. The hard part, which took the nineteenth century, was showing that the resemblances follow rules. Once they did, you could do something astonishing: write down a word that nobody has heard for five or six thousand years, and be able to say which of its sounds were which. This piece covers its two great tools, sound laws and the comparative method, and then four big stories they have been used to tell. Every change here is ordinary speech doing what speech does; no language is ranked. The course in one box Who takes itLinguistics majors, typically in year 3; it is a standard option or core course in programmes shaped by the Linguistic Society of America's guidance and the UK subject benchmark. Classics, archaeology, anthropology and genetics students meet it too. What it is forWorking out how languages are related, how they have changed, and what their ancestors sounded like. Comes afterPhonetics & Phonology (36/21) and Morphology (36/22). Leads toSociolinguistics (36/26), where change is watched while it happens; Languages of the World (36/29), which maps the families; and Computational Linguistics (36/30), which borrows the tree-building methods. What this piece gives youYou will not be able to reconstruct a proto-language. You will see why a "sound law" is not a law of nature, why a cognate table is a kind of detective work, and how linguists, archaeologists and geneticists argue about where Indo-European came from. Easier first reads: The Family Tree of Languages (26/04) and How Language Changes, and Why It Isn't Decay (26/05). Sound change is regular: Grimm's law as a machine Sounds change because people speaking quickly to each other shift their habits by tiny amounts. The surprise is that the shifts are not word by word. When a sound changes in a community, it changes everywhere it occurs in the same surroundings, and so a language's whole vocabulary is rewritten by the same rule at once. The Danish scholar Rasmus Rask noticed the pattern in 1818 and Jacob Grimm set it out in the second edition of his Deutsche Grammatik in 1822. We call the Germanic part Grimm's law: the old voiceless stops p, t, k became the fricatives f, th, h, and the old voiced stops b, d, g became p, t, k. p → f  ·  t → th  ·  k → h    then    b → p  ·  d → t  ·  g → kApplied to every word that had these sounds, in Germanic only. Latin and Greek kept the old sounds. Order matters. In the 1870s a group of young Leipzig scholars, the Neogrammarians, turned the pattern into a working rule: sound change has no exceptions, and an apparent exception means you have not yet found the reason. They were quickly rewarded. Grimm's law seemed to fail for father (Gothic fadar, not *fathar), until Karl Verner showed in 1875 that the pattern depended on where the stress fell in the original word. Sanskrit pitā́ has the stress after the t; that is exactly where Germanic got a voiced d (English th in father is a later development of the same sound). An exception had turned into a second rule. Try it. The tool below applies the rules to nine Latin words (Latin kept the old consonants, so it stands in for the ancestor) and compares the result to the first sound of the English cognate. Then break it by turning the rules the wrong way round. Interactive · Run Grimm's law p t k → f th h b d g → p t k Verner's law Apply voiced rule first Consonants only. Vowels changed under other laws, which is why pedem does not become "fetem" in English. Latin c is written k here. Pairs are standard textbook cognates (Latin cord- is the stem of cor, "heart"). Dating Grimm's law is hard; it happened before written Germanic, probably in the first millennium BCE or earlier. Without JavaScript: with both rules applied in the historical order, all nine words begin with the expected English sound (pater → father, tres → three, duo → two). With the order reversed, duo would become "thuo" because the new t would be shifted a second time. Applying "b d g → p t k" before "p t k → f th h" would shift one sound twice (d → t → th), and two would come out as "thuo". Linguists call the right order of events relative chronology, and you can often establish it with no calendar dates at all. Sound laws apply to one language, in one period, in one set of surroundings. A few changes seem to spread word by word (lexical diffusion, William Wang and later William Labov). Regularity is a very strong working assumption, not a law of nature. The method has a good prediction record. In 1879 Ferdinand de Saussure, from comparing vowels across Indo-European languages, proposed that the ancestor must have had some lost consonant-like sounds. No known language had them. When Hittite was deciphered in the 1910s, Jerzy Kuryłowicz showed in 1927 that Hittite's ḫ appeared in exactly the places Saussure's gaps had been. Nearly half a century before the evidence, the rule had predicted a sound. The Great Vowel Shift: why English spelling lies English spelling is not random. It is a photograph of how the language sounded around 1400, when printing and standard spelling were forming, and the sounds have kept moving ever since. Between roughly 1400 and 1700 the seven long vowels of English rotated, in what scholars call the Great Vowel Shift. In Chaucer's day bite sounded close to modern "beet", house like "hoose", name like "nah-muh", and goose like "gohss". Read English aloud with Chaucer's vowels and the spelling suddenly makes sense. The letters stayed; the sounds slid. The usual picture is a chain shift. The two high vowels (ī and ū) could not rise any further, so they broke into diphthongs, and the vowels below them rose to fill the space. Whether the high vowels pushed the others or the others pulled, and whether it really was one "Shift" rather than several linked changes, is argued. Dates are approximate and differ by region. Step through the stages below. Interactive · Seven words, five centuries Time Simplified transcriptions after standard accounts (Wells 1982; Lass 1999). Stages are approximate and vary by dialect. The chart is schematic: front vowels on the left, back on the right, high at the top; an arrow shows where a diphthong ends. Two letters mark each word (bi = bite, me = meet, ma = meat, na = name, bo = boat, go = goose, ho = house). Without JavaScript: bite went from "beet" to "bite", house from "hoose" to "house", meet rose from "mayt" to "meet", meat rose in two steps and merged with meet by about 1700, name rose from "nah" to "nay", boat from "baw-t" to "boat", and goose from "gohss" to "goose". Regularity again, with a catch. A few words took a different path: great, break and steak kept the older sound and rhyme with late instead of meat. That is not a violation of the rule; it is a reminder that real change spreads through a community, competes with other forms and can be stopped by a social preference. If you meet a "weird" spelling, you are usually looking at a sound that moved after the spelling was fixed. The comparative method: reconstruction from a cognate table Cognates are words in different languages that descend from one word in a shared ancestor: English father and Latin pater; Spanish noche and French nuit. They are not the same as words that merely look alike (Spanish mucho and English much have nothing to do with each other), and not the same as borrowings. The comparative method works in three steps. First, line up cognates and find the sound correspondences, sets of sounds that always match. Second, for each set, work out which single ancestral sound could have turned into all of them by natural, common changes. Third, check that the whole system of changes hangs together. The step that surprises newcomers is the second. You do not take a vote. The Latin cluster -ct- survives in five modern Romance languages in five different shapes, and the shape that wins the vote is not the ancestral one. Try it. This is a cheat in one respect: Latin is written down, so we can check the answer afterwards. In most of the world's families you have no ancestor on paper and the method is all you have. Interactive · Reconstruct the ancestor Four cognate sets. The correspondence is in the shaded sound. Which ancestral sound could every daughter come from? Modelled on standard textbook treatments of Latin -ct- in Romance. The allowed changes (k to t before t, k to a y-glide, k to p, y-glide plus t to "ch") are all common changes in the world's languages. Spelling is used, not phonetic transcription. French nuit and Portuguese noite keep a glide that spelling shows as i. Without JavaScript: the correspondence is Spanish "ch", Italian "tt", French and Portuguese "it", Romanian "pt". Only an ancestral "kt" reaches all of them by common changes (kt → tt; kt → it → "ch"; kt → pt). The majority vote ("it", two languages) is wrong. Latin has noctem, octo, lactem, factum. The method has limits. It cannot see back past roughly six to ten thousand years, when too little of the original vocabulary remains to show a pattern. Claims of older kinship such as "Nostratic" are not accepted by most historical linguists, and vocabulary-percentage dating ("glottochronology") is widely distrusted. Modern work uses Bayesian trees, which come next. Proto-Indo-European and the homeland debate Applied to Indo-European, the comparative method reconstructs a language with a thousand-word dictionary and a grammar: *ph₂tḗr "father", *dwóh₁ "two", *kʷékʷlos "wheel". The asterisk means "reconstructed, never written". The family has at least ten branches: Anatolian (Hittite; extinct), Indo-Iranian, Greek, Armenian, Albanian, Italic (with Latin), Celtic, Germanic, Balto-Slavic and Tocharian (extinct, once spoken in the Tarim Basin in western China). Nearly half the world's people speak a first language from it. Where did its speakers live? Two answers dominate, and genetic evidence has changed the argument since 2015. HypothesisWhere, whenSpread byMain supporters SteppePontic–Caspian grassland, about 4000–3000 BCE (the Yamnaya culture, c. 3300–2600 BCE, is its late form)Pastoralists with wagons, later horses; migrationGimbutas; David Anthony; many historical linguists AnatolianAnatolia, about 7000–6500 BCEThe spread of farmingColin Renfrew; Bouckaert et al. 2012 Each side has real evidence. The steppe case rests on vocabulary: Proto-Indo-European seems to have had words for wheel, axle, yoke, wagon and horse, and wheeled vehicles are not found before roughly 3500 BCE. If the language had split by 7000 BCE, those words would have to be later borrowings or older words that changed meaning. The Anatolian case rests on a simple, elegant mechanism (farming expands, languages travel with it), and on computer-built family trees: Bouckaert and colleagues (2012) put the root at about 8,000 to 9,500 years ago. Linguists have criticised the vocabulary data and methods in those trees. Ancient DNA then changed the picture. Haak et al. and Allentoft et al., both in 2015, found that around 4,500 years ago people with Yamnaya ancestry moved west into Europe on a very large scale, and that central European Corded Ware people were closely related to them. In Britain, Olalde et al. (2018) found that the arrival of the Beaker people from about 2400 BCE was followed by an enormous turnover in ancestry. But DNA is not language: Damgaard et al. (2018) found no steppe ancestry in the Anatolian-speaking Hittite-era people they sampled, which means an Indo-European language arrived in Anatolia without a large migration. Heggarty et al. (2023) built a new tree with ancient languages included and proposed a hybrid: a root around 8,000 years ago, south of the Caucasus, with the steppe as a later centre of spread. In 2025 Lazaridis and colleagues argued from ancient DNA that the earliest ancestors lived between the Caucasus and the lower Volga, around 4400–4000 BCE. Most specialists now treat the steppe region, directly or indirectly, as the source of the main later spread; the Anatolian farmer model survives mostly in a modified form. Treat the exact picture as open. Interactive · Date the root yourself Date of Proto-Indo-European, before any branch split off Each bar is a rough window around the date a published proposal gives, as summarised in the sources listed at the end. A bar lit up means your date falls inside it. This tool does not decide anything: it shows what each date would commit you to. Without JavaScript: the steppe proposal is about 4000–3000 BCE, the Lazaridis et al. proposal 4400–4000 BCE, Heggarty et al. about 6000 BCE, Bouckaert et al. about 7500–6000 BCE, and the Anatolian farmer proposal about 8000–6500 BCE. A date before about 3500 BCE means the shared wheel words must be loans or meaning shifts. A related question is how Indo-European reached South Asia. Narasimhan et al. (2019) found steppe-related ancestry appearing there in the second millennium BCE, consistent with the mainstream linguistic view that Indo-Aryan speakers arrived from the northwest. The view that Indo-European originated in India ("Out of India") is held by some scholars and writers but not by most historical linguists, who point to borrowed Dravidian and Munda features in Vedic Sanskrit and to the shape of the family tree. In India the question is tied to national identity, a reason to keep evidence and politics in separate boxes. Bantu and Austronesian: two great expansions The same tools work outside Europe, and two of the largest language spreads in human history are best read through them. Bantu is a branch of the Niger–Congo family. Its homeland is in the border region of Nigeria and Cameroon, and its speakers spread over several thousand years across central, eastern and southern Africa. Today several hundred languages belong to it (the count depends on what you call a dialect), including Swahili, Zulu, Shona, Kikuyu and Lingala. Austronesian is the family of Malay and Indonesian, Tagalog, Malagasy, Māori, Hawaiian, Fijian and the languages of Taiwan's indigenous peoples: some estimates give 1,257 languages and about 386 million speakers (Wikipedia, citing Ethnologue-based counts). Taiwan is the most likely homeland, with movement outwards starting about 6,000 years ago. ExpansionMilestones (approximate) BantuOrigin in the Nigeria–Cameroon borderlands; the western stream reaches central Angola about 500 BCE; the eastern stream reaches west of Lake Victoria about 500 BCE; one branch reaches KwaZulu-Natal about 300 CE. The causes and routes are debated. AustronesianTaiwan; the Philippines and Indonesia; the Lapita people in the Bismarck Archipelago about 1500 BCE; Fiji, Tonga and Samoa; East Polynesia from about 1025–1120 CE (Society Islands), a wider dispersal about 1190–1290 CE; New Zealand about 1280 CE; Madagascar in the first millennium CE. What makes these two rewarding for a beginner is that the sound laws are visible in names you may already know. Take the word for "person" in five Bantu languages, all from a root *-ntu with a class prefix: Swahili mtu (plural watu), Zulu umuntu (abantu), Shona munhu (vanhu), Kikuyu mũndũ (andũ) and Lingala moto (bato). And across Polynesia, a regular correspondence links the words for "person" and "sacred": GlossMāoriTonganSamoanTahitianHawaiian persontangatatangatatagatata‘atakanaka sacred, forbiddentaputaputaputapukapu Samoan writes the sound ng as g. In Hawaiian, Proto-Polynesian *t became k and *ng became n; Tahitian turned *ng into a glottal stop (‘). So Hawaiian kapu is the same word as Māori tapu, and the English word "taboo" comes from the Tongan form. The same method that links pater and father reaches the Pacific. Madagascar adds a twist: Malagasy is an Austronesian language whose closest relatives are spoken in southern Borneo, thousands of kilometres across the Indian Ocean, although the island's people today combine ancestry from both Southeast Asia and Africa. Language spreading does not prove one population replaced another: in the Bantu case pottery and language do not map one-to-one, and critics call equating them circular. Genetics, archaeology and linguistics each tell part of the story and often disagree. Creoles: new languages in one generation Languages also begin. When people who share no language must work together, as on plantations and in ports, they build a pidgin: a small second language with no native speakers. When children grow up with a pidgin as their main language, it can expand into a creole: a full language with its own grammar and a large vocabulary. Haitian Creole (Kreyòl), the first language of most of Haiti's roughly ten million people, takes its vocabulary mostly from French. Tok Pisin, an official language of Papua New Guinea, is spoken by millions, mostly as a second language, with a vocabulary mostly from English. Others include Jamaican Patois, Krio in Sierra Leone, Nigerian Pidgin (tens of millions of speakers), Seselwa in the Seychelles and Papiamento in the Dutch Caribbean. Creoles are not broken versions of their vocabulary sources. A short sample, glossed in the standard three-line way: Mi go long maket.  (Tok Pisin) 1SG go to market 'I am going to the market.' Yumi go. / Mipela go. 1PL.INCL go / 1PL.EXCL go 'Let's go (you and I) / We are going (not you).' The second line shows something English cannot say in one word. Tok Pisin distinguishes "we including you" (yumi, from "you me") from "we excluding you" (mipela). That distinction is common in the Austronesian languages spoken in Papua New Guinea, a good example of a creole taking its grammar from the language of its speakers rather than from its vocabulary source. Now look at Haitian Creole's verb. Interactive · Build a Haitian Creole sentence Who Marker Verb Forms follow standard descriptions of Haitian Creole (marker te for anterior or past, ap for ongoing action, pral for the future). Real usage has more markers and more nuance. Without JavaScript: in Haitian Creole the verb manje never changes. Li manje "he/she eats"; Li te manje "he/she ate"; Li ap manje "he/she is eating"; Li pral manje "he/she will eat". English has to change the verb itself (eat, eats, ate, eating). Where does this grammar come from? Derek Bickerton's Roots of Language (1981) proposed that creoles formed by children with no adequate model reveal an inborn "language bioprogram". Critics point instead to the African languages that plantation populations already spoke (Claire Lefebvre argued Haitian Creole owes much to Fon and other Gbe languages) or to ordinary mixing from many inputs (Salikoko Mufwene's "feature pool"). John McWhorter argued creoles have simpler grammars than older languages; Michel DeGraff and others reject that. These questions are live. What is not live is whether creoles are real languages: they have native speakers, literature, courts and radio. What a graduate just sees A resemblance between words proves nothing; a regular correspondence between sounds is evidence. You reconstruct the ancestor from the changes that are plausible, not from the form most daughters share. English spelling is a snapshot of c. 1400; the Great Vowel Shift is why "bite" and "meet" are spelled the way they are. Sound laws have reasons behind their exceptions (Verner's law) and sometimes real exceptions (lexical diffusion). Words, genes, pots and peoples spread at different speeds and do not travel as a bundle. There is no "pure" language and none that is a corrupted one. Creoles are simply the newest. The comparative method stops being useful somewhere beyond 6,000 to 10,000 years, and claims past that need extra caution. Where it shows up next Sociolinguistics (36/26): sound change is not only a story about the past; it is driven by social groups, prestige and identity in the present. Languages of the World (36/29) and Phonetics & Phonology (36/21): families, typology and the sound systems that changed. Computational Linguistics (36/30): phylogenetic trees for languages use the same Bayesian tools as biology; Heggarty and Bouckaert used them. In the world: archaeology and ancient-DNA debates, place-name research, legal and educational language policy, the revival of languages such as Māori and Hawaiian, and the documentation of endangered languages before they are lost. If you want to go deeper 📘 Textbook. Lyle Campbell, Historical Linguistics: An Introduction (MIT Press, several editions), is the standard undergraduate text; Terry Crowley and Claire Bowern, An Introduction to Historical Linguistics (Oxford University Press) is another, with many Pacific and Australian examples. 📘 Gentler. David Anthony, The Horse, the Wheel, and Language (2007) makes the steppe case at length; David Reich, Who We Are and How We Got Here (2018) explains the ancient-DNA side; John McWhorter, The Power of Babel (2001) is a lively introduction to language change. 🌐 Free text. The open textbook Essentials of Linguistics (Anderson et al., on the eCampusOntario open library) has chapters on language change; Glottolog (glottolog.org) is a free catalogue of families. 🌐 Free course. MIT OpenCourseWare 24.900 Introduction to Linguistics (Spring 2022) has sessions on language change and on language endangerment, a good free base for this subject. The takeaway Sound change is regular enough that words in related languages answer to each other by rule, and that regularity lets linguists reconstruct languages nobody wrote down. It also tells us where to be modest: it cannot see far past ten thousand years, and the question of where Indo-European began is still being argued with words, bones and DNA, not settled by any one of them. Try this: pick five everyday English words beginning with f and look up their Latin or Greek cousins with p (an etymological dictionary online will do). Then do the same for th and t. Sources & Further Reading Sources. Jones, Third Anniversary Discourse to the Asiatic Society, 1786. Rask 1818; Grimm, Deutsche Grammatik, 2nd ed. 1822; Verner, "Eine Ausnahme der ersten Lautverschiebung," 1875; Osthoff and Brugmann, preface to Morphologische Untersuchungen, 1878; Saussure, Mémoire sur le système primitif des voyelles, 1879; Kuryłowicz 1927. Wang 1969 and Labov on lexical diffusion. Great Vowel Shift: Wells, Accents of English (1982); Lass in The Cambridge History of the English Language vol. 3 (1999). Homeland: Wikipedia, "Proto-Indo-European homeland" (summaries of Gimbutas, Anthony 2007, Renfrew 1987, Bouckaert et al., Science 2012, Haak et al. and Allentoft et al., Nature 2015, Damgaard et al., Science 2018, Heggarty et al., Science 2023, Lazaridis et al. 2025); Olalde et al., Nature 2018; Narasimhan et al., Science 2019. Bantu: Wikipedia, "Bantu expansion". Austronesian: Wikipedia, "Austronesian languages"; Wilmshurst et al., PNAS 2011, and 2008 on rat bones and New Zealand. Creoles: Bickerton 1981; Lefebvre 1998; Mufwene 2001; McWhorter 2001; DeGraff 2003. The wheel's date (around 3500 BCE) and the number of Indo-European branches are textbook figures. Calculations on this page (the sound-change tables, the vowel chart and the cognate evaluator) were run in node. This piece describes research; dates are approximate and several are disputed. 📘 Campbell, Historical Linguistics: An Introduction 📘 Anthony, The Horse, the Wheel, and Language 🌐 Glottolog Step 8. Sociolinguistics: Nobody Speaks Just the Language --------------------------------------------------------- Why: Dialect, class, identity and language policy. Every speaker speaks a particular variety, to a particular person, in a particular mood. Sociolinguistics measures that variation, finds the rules in it, and asks who gets to call what "correct." In 1962 a young linguist walked into three New York department stores, one expensive, one middling and one cheap, and asked shop assistants where to find something he knew was on the fourth floor. "Fourth floor," they said. Then he said "Excuse me?" and they said it again, more carefully. He was not shopping. He was counting how often each person pronounced the r in fourth and floor. The linguist was William Labov, and the stunt was a way of asking an old question in a new way: does the way people talk follow social patterns, and can you measure them? It does, and you can. Nobody in those stores was speaking carelessly. Every one of them was following a rule, one that depended on their job, the store, and how closely they were watching their own words. The graduate of this course keeps a shift in attitude. A dialect is not a damaged version of a standard, and an accent is not a lack of one. Variation is the normal state of any language, and it carries social meaning that speakers read in a fraction of a second. The course in one box Who takes itLinguistics majors, usually in year 3, and sociology, anthropology, education and language-teaching students. It is a core area in the LSA's guidance and in UK linguistics degrees. Some programmes teach it as variationist work with recordings and statistics; others lean on anthropology and policy. What it is forDescribing how language varies with place, class, age, setting and identity, and how societies manage the languages they contain. Comes afterPhonetics & Phonology (36/21), because most variables are sounds. Leads toHistorical Linguistics (36/25), Languages of the World (36/29), Language Acquisition (36/27), and, outside the major, public policy and education. What this piece gives youYou will not be able to run a sociolinguistic survey. You will know six ideas that change how you hear speech, and some real language policies. Easier first reads: How Language Changes, and Why It Isn't Decay (26/05) and The Family Tree of Languages (26/04). Variables: Labov and the sounds of a city A sociolinguistic variable is something that can be said in two or more ways without changing the meaning: -ing or -in', an r after a vowel or none, I don't have any or I don't have none. Each way is a variant. Labov's idea was that you do not ask whether a speaker "has" a feature. You count how often, in which setting, among which people. The department store study (carried out in 1962, published 1966) compared Saks (upmarket), Macy's (middle) and S. Klein (discount). He gathered a little over 250 short interviews. The pattern was the one he predicted: the expensive store's staff used the prestigious r most, the discount store's least, and Macy's in between. He also found that within each store, the repeated "careful" version had more r. Speakers shift towards the prestige form when they pay attention. In a wider survey of New York, Labov found something more striking. The lower middle class, when reading word lists aloud, overshot: they used more of the prestigious form than the upper middle class did. He called this hypercorrection: the group most anxious about status over-applies the rule. Interactive · Style, class and the crossover Speaking style (from casual to reading minimal pairs) Schematic: invented numbers shaped like the pattern in Labov's New York work (the lower middle class crossing above the upper middle class in the most formal styles). Not his data. LC lower class, WC working class, LM lower middle class, UM upper middle class. Without JavaScript: in casual speech the upper middle class pronounces the most r and the lower class the least. In word lists and minimal pairs, the lower middle class overtakes the upper middle class. Labov's other early study, on Martha's Vineyard, Massachusetts (1963), showed that a sound change can carry identity. Islanders, particularly fishing families wanting to stay, raised the starting vowel of words like right and house more than those who planned to leave, as the island filled with summer visitors. The change was social: a way of saying I belong here. Variation also tracks age. Comparing generations at one time ("apparent time") lets researchers see change happening, since younger speakers lead most sound changes. Studies of women and men across many cities have often found women leading changes towards the standard, and sometimes changes away from it, a pattern that is robust but not universal, and which is explained differently by different researchers. Dialects, continua and who gets called a language "A language is a dialect with an army and a navy." The line is usually credited to the Yiddish scholar Max Weinreich, who in 1945 said that a listener had given it to him. It makes a real point. Linguistically, there is no sharp border between a language and a dialect. The popular test is mutual intelligibility, but it fails both ways. Swedish, Norwegian and Danish speakers understand one another to a good degree, and they are three standard languages. Mandarin and Cantonese are called dialects of Chinese, but their speakers cannot understand each other's speech without study. In both cases the label follows politics, history and writing, not sound. The reason is the dialect continuum. Walk from village to village in a rural region, in Germany and the Netherlands, along the Romance-speaking regions of Europe, or across northern India, and each village sounds only slightly different from the next, yet the two ends of the walk may not understand each other. Dialectologists draw isoglosses, lines on a map where one variant gives way to another. When many lines bunch together you have a dialect boundary, though it is rarely clean. Interactive · A dialect map of twelve invented villages Village Compare with The villages, words and sounds are invented for illustration. Dark dots use the eastern variant of the features you chose; dashed lines are isoglosses. Without JavaScript: in the invented valley, villages next to each other differ in at most one of five features, while villages 1 and 12 differ in all five; three of the five lines bunch around villages 5 to 8. Notice what the tool lets you do. No single line splits the valley, and the three lines at the ridge give the best case for a boundary. Where you draw "two dialects" depends on how many features you require. The same problem sits under the count of the world's languages. Ethnologue's recent editions list a little over 7,100 living languages, and Glottolog and other catalogues give different totals, because each must decide where a language ends and a dialect begins. Dialects and creoles are full languages "Standard" languages are dialects that won: the dialect of a capital, a court or a printing trade, then codified in dictionaries and schools. It is not more logical or more expressive. It is the one with institutions behind it. Linguists use "standard" and "non-standard" for that reason, and avoid "correct" and "broken." Labov showed the point with African American English (AAE) in a 1969 paper on the English copula, the verb to be. Standard English can contract is to 's and are to 're: He's nice. AAE can delete it: He nice. Labov noticed that deletion happens exactly where Standard English can contract, and nowhere else. You can say He's tall, but not That's what he's at the end of a sentence. AAE likewise says He tall, but never That's what he. This is a precise rule, shared with Standard English, applied one step further. Interactive · Contract or delete? After Labov's 1969 analysis. The sentences are textbook-style; real speech varies, and AAE is a family of varieties with its own history and regional forms. Without JavaScript: He is nice contracts (He's nice) and deletes (He nice). I know where he is does neither: neither *I know where he's nor *I know where he is possible. AAE has other precise features. He be working (habitual be) means he works regularly, while He working means he is working now: a distinction Standard English marks with adverbs. Lisa Green's African American English: A Linguistic Introduction (2002) is the standard description. Creoles are the clearest case of languages that outsiders dismissed. A pidgin is a simplified contact language used between groups with no shared language, and has no native speakers. When children grow up speaking it as a first language, it becomes a creole, with a full grammar. Derek Bickerton argued in the 1980s that children's own inborn capacity creates the grammar, a claim that was widely debated; John McWhorter and others stress features that creoles share because of their young age. Creoles are official languages in several places: Haitian Creole is, with French, an official language of Haiti under the 1987 constitution, and Tok Pisin is among the official languages of Papua New Guinea alongside English and Hiri Motu. Mwen te manje1SG PST eat"I ate." (Haitian Creole; the past marker comes before the verb) Yumi go1PL.INCL go"Let's go (you and I)." (Tok Pisin) Mipela i go1PL.EXCL PM go"We (not you) are going." (Tok Pisin) Tok Pisin has a distinction that English lacks: yumi ("we, including you") against mipela ("we, excluding you"). Many other languages do the same, including Quechua and Tamil. Code-switching: a skill, not a failure Bilingual speakers move between languages within a conversation, or even within a sentence. A common view, held for decades, was that this mixture showed confusion or a lack of full command of either language. Shana Poplack's study of Puerto Rican speakers in New York, in a paper titled with a sentence that does what it describes ("Sometimes I'll start a sentence in Spanish y termino en español," 1980), found the opposite. The most fluent bilinguals switched within sentences most smoothly, and they switched where the grammars of both languages line up. Carol Myers-Scotton's Matrix Language Frame (1993) gave another model: one language supplies the grammatical frame, and the other fills in. Why switch? John Gumperz (1982) listed functions: quoting, choosing an addressee, stressing a point by repeating it, marking a change of topic or activity, and signalling shared identity. Jan-Petter Blom and Gumperz (1972), in Hemnes, Norway, found that people shifted between the local dialect and the standard when the situation changed (an office, a street) and also to mark a change of topic with the same people. Interactive · Why did they switch? Pick a function for each utterance. Invented utterances in the style of reported bilingual speech. Real switches often serve several purposes at once. Without JavaScript: the Spanish quotation inside an English story is a quotation; the repeated Hindi warning is emphasis; the Welsh greeting before English business marks a change of activity; the child-then-visitor switch picks an addressee; the Indonesian-English "bro" marks shared identity. Switching differs from diglossia, Charles Ferguson's 1959 term for two varieties with different jobs in one society, such as a written standard and a spoken vernacular in the Arabic world, German-speaking Switzerland, Haiti and the Greek of that time. In diglossia the choice is set by the occasion; in code-switching, speakers are using both languages by choice at once. Language and identity: who we sound like Why would anyone pronounce words in a way their teachers marked wrong? Because it works. Peter Trudgill's 1972 study of Norwich, England, found that many working-class men claimed in tests to use the standard forms less than they did, and many women claimed more. He called the pull of the non-standard form covert prestige: it signals toughness, solidarity and belonging, and not wanting to sound like "them." Penelope Eckert's study of a Detroit-area high school (published 2000) found that students' vowels followed their social category: the "jocks," who looked to the wider world, and the "burnouts," who rooted themselves locally, sounded different, and the pattern tracked how they saw themselves, not their parents' income alone. Sound carries consequences too. In a 1999 study by Thomas Purnell, William Idsardi and John Baugh, Baugh phoned landlords about advertised flats in several Bay Area neighbourhoods. He used three voices (standard American English, African American English and Chicano English). Callers heard as African American or Chicano were less often offered appointments, and the gap varied with the neighbourhood. Rosina Lippi-Green's English with an Accent (1997) names the thing behind it: a standard language ideology, the belief that one variety is simply better, which treats judgments about speech as if they were judgments about people. The implication is not that standards should vanish. A shared written standard has real uses: it lets strangers read each other, and it opens doors that many people have to go through. The sociolinguist's claim is narrower: the standard's advantages are social, so teaching it is a good idea, and calling everything else wrong is not. Language policy: who decides what is official Governments choose languages for schools, courts and offices, whether they intend to or not. Four cases show the range. The tool below sets out the written rule, the dated numbers, and where tension remains. Interactive · Four countries, four answers Constitutional and census facts as of the latest figures found in 2026; they change. This describes policy and is not legal advice. Without JavaScript: India's Union official language is Hindi in Devanagari, with English continued under the Official Languages Act 1963, and 22 scheduled languages. Paraguay has two official languages, Spanish and Guaraní (1992). Spain's Castilian is official nationally and Catalan, Galician, Basque and Aranese are co-official in their regions. South Africa has 12 official languages since the 2023 addition of South African Sign Language. The cases make an argument among themselves. India's states were redrawn largely along linguistic lines in the 1950s because a single shared language was not available, and attempts to push Hindi nationally have met strong resistance, especially in Tamil Nadu in the 1960s. Paraguay is unusual: the indigenous language is spoken by a large share of non-Indigenous citizens as well, and mixing with Spanish (Jopara) is everyday speech. Spain's model gives regions the power to co-officialise their languages, and disputes over the language of schooling are argued in courts and parliaments by people who think they are protecting a language and others who think they are protecting individual choice. South Africa's constitution gives eleven spoken languages parity on paper; critics point out that English dominates government and business in practice, and defenders say the framework still protects languages that had no official status before 1994. These are values questions as well as linguistic ones. The linguist's contribution is the evidence: which languages are dying (UNESCO and Ethnologue both count a large share of the world's languages as endangered), what children learn best in (research on early mother-tongue teaching generally favours it), and what a rule costs to apply. The decision is the community's. What a graduate just sees Variation has rules: count how often, among whom, in what setting. People shift towards prestige forms when they pay attention, and anxious groups overshoot. A dialect continuum has no clean border; "language" versus "dialect" is a social decision. AAE, like every variety, follows precise rules, and creoles are complete languages. Code-switching is a skill that fluent bilinguals use, with grammar and with purposes. Non-standard speech can carry covert prestige and belonging. Judging speech is often a way of judging people, and the standard's advantage is social, not logical. Language policy is a choice about whose language the state uses. Where it shows up next Historical Linguistics (36/25): sound changes begin as variation inside a community, as on Martha's Vineyard. Languages of the World (36/29): how many languages there are, how endangered, and why counts differ. Language Acquisition (36/27): children acquire the variety they hear around them, including its social meaning. Computational Linguistics (36/30): speech recognition that works for some accents far better than for others. In the world: classroom practice, courtroom interpreting, hiring and housing bias, and language revival from Welsh to Māori to Hawaiian. If you want to go deeper 📘 Textbook. Miriam Meyerhoff, Introducing Sociolinguistics (Routledge, 3rd edition 2018), or Janet Holmes and Nick Wilson, An Introduction to Sociolinguistics (Routledge, 5th edition 2017). Check for a newer edition. 📘 Gentler. Rosina Lippi-Green, English with an Accent (2nd edition 2012); John McWhorter, The Power of Babel (2001). 🌐 Free text. Labov's "The Logic of Nonstandard English" (1969) is widely available online, and the open textbook Essentials of Linguistics has a chapter on language variation. 🌐 Free course. MIT OpenCourseWare 24.900 Introduction to Linguistics (Spring 2022) includes units on dialects and on historical change. The takeaway Variation is not noise on top of a language; it is the language in use. It follows rules, it carries identity, and it is judged socially. Dialects and creoles are complete languages, bilingual mixing is a skill, and official policy is a choice about whose speech counts. Try this: for one day, listen for a single variable (say -ing versus -in') in two people you know and in yourself at work versus at home. Notice when it changes. Sources & Further Reading Sources. Labov, "The social motivation of a sound change," Word 1963; The Social Stratification of English in New York City (1966; department stores 1962, sample of 264 interviews as reported in secondary summaries, individual store percentages not reproduced here); Labov, "Contraction, deletion, and inherent variability of the English copula," Language 1969, and "The logic of nonstandard English" (1969); Green, African American English (2002); Weinreich, YIVO Bleter 1945; Bickerton 1981 and McWhorter 2001 on creoles; Poplack, Linguistics 1980; Myers-Scotton 1993; Gumperz, Discourse Strategies (1982); Blom and Gumperz 1972; Ferguson, "Diglossia," Word 1959; Trudgill, Language in Society 1972; Eckert, Linguistic Variation as Social Practice (2000); Purnell, Idsardi and Baugh, Journal of Language and Social Psychology 18 (1999), 10–30; Lippi-Green 1997. Policy facts: Constitution of India, Art. 343 and the Eighth Schedule (22 languages), Official Languages Act 1963, Census of India 2011 (121 languages, 1,369 rationalised mother tongues, Hindi 43.63% as first language); Paraguay Constitution 1992, Art. 140 and the 2022 census (monolingual Guaraní 28% in 2002, 12% in 2022); Spain Constitution 1978, Art. 3; South Africa Constitution 1996 s. 6, Statistics South Africa Census 2022 (home language shares) and the 2023 vote adding South African Sign Language; Haiti Constitution 1987; Papua New Guinea (Tok Pisin). Language counts: Ethnologue and Glottolog, latest editions, about 7,100. The crossover chart, village distances and rule outputs were computed in node. This piece describes how people speak and does not rank any variety. 📘 Meyerhoff, Introducing Sociolinguistics 📘 Lippi-Green, English with an Accent 🌐 Glottolog Step 9. Translation & World Literature: Who Carries a Book Across ----------------------------------------------------------------- Why: Translation, and literature across languages. Every book you read from another language is someone's choices. And whether it crossed the border at all was decided by people, prizes and money. Think of the great books you have read: the Odyssey, War and Peace, the Thousand and One Nights, One Hundred Years of Solitude, a Murakami or a Han Kang. For most readers of English, nearly every one of those was written in another language. You did not read Homer. You read a translator's decisions about Homer, made in a particular century for a particular audience, and you probably never saw the translator's name on the cover. Nobody who studies literature escapes this. Whenever a book crosses a border, someone has to choose what travels. And whether it travels at all is decided by people and money well before any reader sees it. This piece is about both: what a translation keeps and loses, and what we mean when we talk about "world literature" at all. The course in one box Who takes itComparative-literature, English and modern-languages students, usually in year 3, and translation-studies programmes. Courses with names like "World Literature" or "Masterpieces of World Literature" are taught in many countries, and their reading lists differ a lot. What it is forReading a book from another culture without pretending the translation is the original, and seeing how books, prizes and publishers shape what any of us thinks of as "great". Comes afterLiterary Theory & Criticism (35/34), which supplies the questions about power and canon this piece applies. Leads toIt sits at the end of the Literature path. It connects back to Epic, Myth & the Ancient Literatures (35/30) and Modern, Postcolonial & Contemporary Literature (35/33). What this piece gives youYou will not be able to translate a poem from a language you don't read. You will be able to look at two translations of one line and say what each chose, and to read a prize list as evidence about a system. Easier first reads in the library: Where English Words Come From (26/03) and World Cinema: A Starter Map (10/09). What translation keeps, and what it loses Translation looks like swapping each word for its twin. It is closer to rebuilding a house in a different country: same plan, different bricks, different weather. Roman Jakobson (1959) separated three jobs: rewording inside a language, translating between languages, and turning words into something non-verbal. The middle one is where meaning, sound, rhythm, tone, culture and allusion all pull in different directions. You can keep some. You can almost never keep all of them at once. Short poems show it best. Take a famous Chinese quatrain, "Quiet Night Thought" (Jìng yè sī) by Li Bai (701–762, Tang dynasty). It has twenty characters. The English below is my own literal rendering of the public-domain original. Chinese verse has no tense markers and often leaves out who is doing the looking. English demands both. Even the original is not fixed: editions from the Song to the Qing dynasties differ by a character or two, for example "bright moon" versus "mountain moon" in the third line. LineCharactersWord by word 1床前明月光bed · front · bright moon · light 2疑是地上霜suspect · is · ground · on · frost 3舉頭望明月raise · head · gaze · bright moon 4低頭思故鄉lower · head · think of · old home To make an English poem out of that, a translator must decide who "I" is, whether the doubt in line 2 is a puzzled "I wonder" or a plain "it seems", and how warm to make the verb in line 4. Build your own and see how many defensible versions come out of twenty characters. Interactive · Build a translation of Li Bai Who is looking? Line 2: how sure is the speaker? Line 4: the verb Line 4: the place Eighty-one versions, from four small choices, and each is a faithful translation of the same twenty characters. This is why translators say they are never done, and why there are many English versions of the same poem. Foreignising or domesticating: bring the author or bring the reader In 1813 the German theologian Friedrich Schleiermacher gave a lecture, "On the Different Methods of Translating", saying there are basically two roads. Either leave the author in peace and move the reader toward him, or leave the reader in peace and move the author toward the reader. Nearly two centuries later Lawrence Venuti gave the two roads names, in The Translator's Invisibility (1995): domesticating (smooth, fluent English that hides the translation) and foreignising (keeping some strangeness so the reader remembers the book came from elsewhere). Venuti argued that English-language publishing overwhelmingly rewards the first, and that this makes translators invisible and other cultures easier to flatten. Defenders of fluency answer that most readers want a book, not a lesson, and that a smooth English sentence can be a faithful one. Eugene Nida, translating the Bible, called a similar split formal and dynamic equivalence (1964). Seeing it in practice takes one line. The Odyssey starts with a Greek word, polytropos, that describes its hero: literally "of many turns". Here are four public-domain English versions of the opening, from four centuries. Press each one. Interactive · One Greek word, four centuries Ἄνδρα μοι ἔννεπε, Μοῦσα, πολύτροπον, ὃς μάλα πολλὰ Translators: George Chapman (about 1615), Alexander Pope (1725–26), William Cowper (1791), Samuel Butler (1900). All public domain. The notes on what each does are mine. Chapman keeps the picture of turning and winding. Pope, in rhymed couplets, turns it into worldly skill. Butler turns it into plain prose. A recent version, Emily Wilson's (2017), made headlines by calling the hero "complicated", a choice that is bold exactly because it is so ordinary. None of them is wrong. Each answered a different question about what an English reader of its time could bear. The same goes for the other great texts from outside Europe. The Tale of Genji (early eleventh century) has been retranslated again and again: Arthur Waley (1925–33), Edward Seidensticker (1976), Royall Tyler (2001) and Dennis Washburn (2015), among others. The original mostly names its characters by rank or place instead of by name, and each translator has had to decide how much of that to keep. Goethe's "world literature" The phrase Weltliteratur is Goethe's. On 31 January 1827 he told his young assistant Johann Peter Eckermann that he had been reading a Chinese novel, that national literature no longer meant much, and that the epoch of world literature was at hand. (We know it from Eckermann's book Conversations with Goethe, first published in the 1830s, so these are his words as Eckermann remembered them.) What lit Goethe's idea was translation. His West-östlicher Divan (1819) was inspired by the Persian poet Hafez, whom he read in a German version, and the Chinese novel he mentions reached him in a European translation too. Twenty-one years later Marx and Engels used the same phrase for a different purpose. Describing how markets were remaking the world, the Communist Manifesto (1848) says "from the numerous national and local literatures, there arises a world literature" (Samuel Moore's 1888 English version). For Goethe it was a conversation. For Marx and Engels it was a by-product of trade. The two views are still the two poles of the debate. Today's main voices disagree like this. David Damrosch (What Is World Literature?, 2003) says world literature is not a list but a mode of circulation: works that gain life by being read outside their culture of origin, and change in the process. Pascale Casanova (The World Republic of Letters, French 1999) argues that literature has its own unequal geography, with prestigious "capitals" (she meant Paris above all) that consecrate writers from the margins. Franco Moretti asks us to see it as one world system with a centre and edges. Gayatri Spivak and Emily Apter (Against World Literature, 2013) worry that "world literature" erases the untranslatable and the languages that rarely get translated. And scholars point out that Europe did not invent world literature: Sanskrit, Persian, Arabic and Classical Chinese each carried a literature across a vast area long before the word existed. How a book travels, and who decides Between a finished novel in Korean and a reader in Kansas stand a series of doors, and each is opened by someone with taste, budget and risk. A local publisher must bring it out. An agent or a scout must notice it. A translator must be found and persuaded. A publisher must buy it, often with help from a translation grant. Reviewers, bookshops and prizes must then give it a place to sit. Lose it at any door and it does not arrive. The numbers show how narrow the road is. The "three percent" figure, first established by the bibliographic firm Bowker and made famous by Chad Post's Three Percent project at the University of Rochester, says only about 3% of books published in the US are translations. For literary fiction and poetry the share is lower, roughly 0.7 to 1% by most counts, and sources disagree on method (the database counts only titles never before in English; Post himself said in 2014 that for fiction and poetry it sits at about three percent, while noting it depends on what counts as a book). The point holds either way: very little of the world's writing reaches English, and a small number of publishers in a few cities choose which. That is the setting for the toy below. The five door-percentages are illustrative guesses, not measurements. Move them and watch what happens to a thousand books. Notice how a small change at one door ripples through the whole product. Interactive · A thousand books, five doors (illustrative) Start point Agent and a grant Small language, no agent One real-world nudge is worth knowing. The International Booker Prize, since 2016, is awarded to a single translated novel or story collection each year, and the £50,000 prize is split equally between author and translator. The split is a public statement that the translator did half the work. The Nobel and the world map The Nobel Prize in Literature has been given since 1901. Per the Nobel Foundation's own tally as of the 2026 prize (announced October 2026, to the Canadian poet and translator of Greek Anne Carson), the prize has been awarded 119 times to 123 laureates, 19 of them women. It is the world's most famous literary prize, so what it shows is worth reading as evidence about the system, and not as a ranking of writers. Slide the year below to see where the laureates were born, region by region. The regions are crude, and birthplace is not literary home: Rudyard Kipling (born in Bombay) counts as Asia here, and Albert Camus (born in French Algeria) counts as Africa, although both are filed in the literature of the language they wrote in. Interactive · Where the laureates were born Prizes up to the year Data: Nobel Foundation API (nobelprize.org), fetched November 2026; birth country as today's country; Gurnah (born in Zanzibar) counted as Africa. By 2026, 83 of the 123 laureates, about two in three, were born in Europe, 13 in North America, 11 in Asia, 9 in Latin America and the Caribbean, and 7 in Africa. The first Asian laureate was Rabindranath Tagore in 1913, and the prize citation praised his poetry as expressed in his own English words (Gitanjali, 1912, with a preface by W. B. Yeats). Gabriela Mistral of Chile (1945) was the first Latin American. Wole Soyinka (1986) was the first from sub-Saharan Africa, and Naguib Mahfouz (1988) the first writing in Arabic. Recent winners range from Abdulrazak Gurnah (2021, born in Zanzibar) to Han Kang (2024, South Korea), but the early pattern is long. Whether the prize reflects the world's literature or one academy's reading of it is exactly the argument between Damrosch, Casanova and Apter. Reading in translation well None of this means translations are second-best. It means you read them differently. A few habits cost nothing. Look for the translator's name and the year, and read their note. It will say what they chose. Expect retranslation. Languages age. A version from 1900 is a document of 1900 as well as of the original; read more than one for books you love. Compare one page. Put two translations side by side and ask what each made easy to see and what it hid. Watch for the hard cases where translators must choose: forms of address (the French tu and vous), Japanese honorifics, puns, names that mean something, rhyme. Notice smoothness. If a book from a very different culture reads like a native English novel, ask what was ironed out. Read the original's own world. A translation is one road in; a history, a map and a second book from the same literature are others. The writer Ngũgĩ wa Thiong'o made the opposite point in Decolonising the Mind (1986): he stopped writing in English and wrote in Gikuyu, then translated himself into English, so that the original and the "translation" changed places. If language is a way of holding the world, the direction of translation is never neutral. What a graduate just sees A translation is a set of choices, and there is never just one right set. The smooth reading is the one that hides the choices. Each generation retranslates because its own English has changed. Whether a book travels depends on doors held by agents, publishers, grants and prizes, not only on quality. "World literature" means two things at once: a conversation across languages and a market with centres and edges. A prize list is evidence about the system that produced it. Every canon is also a map of who had access to the printing press and the translator. Being read in the original is a different experience, not a better one, and no one reads in all the languages. Where it shows up next Epic, Myth & the Ancient Literatures (35/30): every ancient text you read is already a translation, so everything here applies. Modern, Postcolonial & Contemporary Literature (35/33): the choice of language, from Achebe's English to Ngũgĩ's Gikuyu, is a literary and political act. Language & Words (26): the same questions about meaning and borrowing, from the linguist's side. In the world: publishing, subtitling, diplomacy and machine translation all face the same trade-offs; today's automatic tools are fluent by default, which is Venuti's domesticating road at scale. If you want to go deeper 📘 Textbook. David Damrosch, What Is World Literature? (Princeton University Press, 2003). The standard modern statement of the idea, widely assigned in comparative-literature courses. 📘 Textbook. Lawrence Venuti, The Translator's Invisibility: A History of Translation (Routledge, 1995; 3rd edition 2018). The classic on fluency and foreignising; or Pascale Casanova, The World Republic of Letters (Harvard University Press, translated by M. B. DeBevoise), for the geography of prestige. 📘 Gentler. David Bellos, Is That a Fish in Your Ear? Translation and the Meaning of Everything (2011). Funny, practical and written for general readers. 🌐 Free course. HarvardX, Masterpieces of World Literature: a free online course on world classics from several traditions, led by Harvard faculty. The takeaway Every translation is a bundle of choices about what to keep, and the smooth version hides them. Books also cross borders through doors held by publishers, prizes and money, so "world literature" is both a real conversation between cultures and an unequal market. Read translations as the work of a named person at a given time, and treat a prize list as a clue to how the system works. Try this: this week, find two translations of one famous opening line, whether a Bible verse, a first line of the Odyssey or of a poem you like, and write down in one sentence what each translator chose that the other did not. Sources & Further Reading Sources. Public-domain translations quoted from Project Gutenberg: Chapman (ebook 48895), Pope (3160), Cowper (24269), Butler (1727); the Greek opening is Odyssey 1.1. Communist Manifesto, Moore translation 1888 (Gutenberg ebook 61). Li Bai's original text and the Song-to-Qing textual variants: Wikipedia, "Quiet Night Thought" (citing Minford and Lau, Classical Chinese Literature, 2002); the English glosses are my own. Goethe, 31 January 1827: Eckermann's Gespräche mit Goethe (first published in the 1830s), as summarised by Heidelberg and Lapham's Quarterly sources; Eckermann's wording is a reconstruction. Translation share of US books: Three Percent / Translation Database, University of Rochester. International Booker split: The Booker Prizes (2016 onwards). Emily Wilson's "complicated" first line: reported in coverage of her 2017 Odyssey. Nobel counts, laureates, birth countries and the 2026 prize: nobelprize.org and its API (119 prizes, 123 laureates, 19 women, fetched November 2026). Regional counts (Europe 83, North America 13, Asia 11, Latin America and Caribbean 9, Africa 7) were computed from that data. Schleiermacher 1813, Venuti 1995, Nida 1964, Jakobson 1959, Damrosch 2003, Casanova 1999, Apter 2013: standard references in translation studies and comparative literature. 🌐 HarvardX: Masterpieces of World Literature 🌐 The Nobel Prize in Literature (nobelprize.org) 📘 Damrosch, What Is World Literature? 📘 Bellos, Is That a Fish in Your Ear? Step 10. Language Acquisition: How Children Build a Language ------------------------------------------------------------ Why: How children learn language, and what that says about the mind. Every typical child does something no adult can easily match, and the argument over how is one of the oldest in the cognitive sciences. A two-year-old says "I runned to the door." She has never heard an adult say "runned". Nobody told her the rule for the past tense. She worked it out, from the speech around her, and then applied it a little too widely. Within two years she will say things like "I wonder whether the dog that I saw is still outside", a sentence with a question inside a clause inside a clause, which she has also never been taught. Every typical child, in every country, in every language and in sign, does this. No other animal does, and no adult learner finds it easy to match. The question of how is one of the oldest quarrels in the cognitive sciences, and it is more than a quarrel about children: it is about what a mind starts with and what it takes from the world. This piece goes past the basic timeline in How Children Learn Language (26/06) to the experiments and the arguments. The course in one box Who takes itLinguistics majors in year 4, and psychology, education and speech-language-pathology students, who often meet it earlier; it sits at the border between linguistics and developmental psychology. What it is forUnderstanding how children go from babbling to grammar, what they bring to the job, and what the evidence from deprivation, bilingualism and sign languages says about it. Comes afterSyntax (36/23) and Developmental Psychology (35/24). Leads toPsycholinguistics & the Brain (36/28) and Computational Linguistics (36/30), where learning models meet language models. What this piece gives youYou will not be able to assess a child. You will know the main findings, the two big theories and what each says about the same data, and which famous stories need care. Easier first reads: How Children Learn Language (26/06) and Learning a Language as an Adult (26/07). Not clinical advice. This describes research on typical development. Children vary a great deal. If you are concerned about a child's hearing or speech, a paediatrician or speech-language pathologist is the right person to ask. The first year: statistics before words Speech does not come with spaces. In a continuous stream of sound, a baby has to find where one word ends and the next begins, in a language she does not yet know. In 1996 Jenny Saffran, Richard Aslin and Elissa Newport tested a clever possibility. They played 8-month-old babies about two minutes of a made-up language, a monotone stream built from four nonsense words (tupiro, golabu, bidaku, padoti) in random order, with no pauses and no stress. Within a word, one syllable always follows another (tu is always followed by pi); across a word boundary, the next syllable is one of three possibilities. Afterwards the babies listened longer to new orders that mixed up syllables from different words than to the "words" themselves. They had tracked transitional probabilities: how likely one syllable is to follow another. TP(B | A) = number of times A is followed by B ÷ number of times A occursHigh TP inside a word; a dip at the boundary. The dips are where the words break. This finding has held up well, and it also works for non-speech tones and in other species (Hauser, Newport and Aslin found it in cotton-top tamarins in 2001). It does not mean this is the only cue: real infants also use stress patterns and pauses. Try the learner. Interactive · Find the words without knowing any Words heard in the stream A dip below this probability marks a boundary A simulated stream from the four words used by Saffran et al. (1996), in a fixed pseudo-random order with no word repeated twice in a row. Green bars mark a correct boundary the learner found, red a boundary it hit in the wrong place, grey a boundary it missed. The babies' own results were about listening times, not a count like this. Without JavaScript: inside each word the transitional probability is 100%; at a word boundary it is much lower (in the simulation, roughly a third to two fifths). With enough listening, a rule "break where it dips" finds almost every word. Babies also narrow their ears (the 9-month Mandarin experiment is in 26/06). Sound, rhythm and statistics are the first year's job. The first word usually arrives at around 12 months; by then the child has been learning her language for a year without producing any of it. First words and the rules children invent After the first words (typically around 12 months), vocabulary grows slowly, then faster. By 18 to 24 months many children combine two words ("more milk", "daddy go"). Normal range is wide: a child with 50 words and a child with 300 at two years can both be developing normally, and late talkers often catch up, though not always. Languages differ in what is easy. Turkish words are built by stacking suffixes, as in ev-ler-im-de, and children learning Turkish tend to use these suffixes correctly remarkably early, an observation of Dan Slobin's cross-linguistic work. ev-ler-im-de  (Turkish) house-PL-1SG.POSS-LOC 'in my houses' The error that captured everyone's attention is overregularization. For a while a child says went, then later goed, then went again. The child did not hear "goed"; she extracted the -ed rule (Jean Berko's 1958 "wug test" shows children apply it to new words) and applied it to a verb that breaks it. But the picture is easy to exaggerate. In their 1992 monograph, Gary Marcus and colleagues analysed 11,521 irregular past-tense uses from children's recorded speech and found overregularization in a median of just 2.5% of them. Later analyses give a higher average (4.2% in one dense corpus, with one child at 24%). Maratsos (2000) argued that averages hide the fact that rare verbs are far more error-prone. Children, in short, get irregular verbs right most of the time, and errors cluster on the less common ones. A simple "memory plus rule" model reproduces this pattern. For each irregular verb, the child stores the past form once heard; the more often she hears it, the stronger the memory. If retrieval fails, the rule fills in. Steven Pinker and Alan Prince's "words and rules" account (1988) is built on this idea; connectionist models (Rumelhart and McClelland, 1986) argue that one network can do both jobs. The tool below is a toy of the first kind, with invented frequencies. Interactive · A memory-plus-rule toy Child's age (the rule is learned at 24 months here) Illustrative model, not data. Chance of retrieving a past form = 1 − exp(−hearings ÷ 20); hearings = verb frequency × months of listening since 12. Frequencies are invented. The point is the shape: errors appear when the rule arrives, fall as memories strengthen, and stay highest for rare verbs. Without JavaScript: in the toy, errors begin at the age the rule is learned, fall steadily afterwards, and are highest for the rarest verbs; the overall average is a few per cent, as in the real data. Poverty of the stimulus versus learning from use How can a child learn grammar from input that is messy, incomplete and never labelled? Noam Chomsky's answer (from the late 1950s) is that she cannot do it from experience alone: the data underdetermine the grammar, so children must start with some built-in structure, which he called universal grammar. This is the poverty-of-the-stimulus argument. The standard example is question formation. A child hears "The man is happy" and "Is the man happy?" and could conclude: move the first "is" to the front. That rule fits every simple sentence perfectly and is wrong for "The man who is tall is happy". Interactive · Which rule did the child learn? Simple questions heard Questions with a clause inside the subject heard The two rules are the textbook ones. Hearing counts are for you to play with: how many complex questions real children hear is one of the points in dispute. Without JavaScript: for "The man who is tall is happy", the linear rule gives "Is the man who tall is happy?", which no English speaker says; the structural rule gives "Is the man who is tall happy?". The two rules agree on every simple sentence. The tool shows the argument: with only simple questions the two rules are indistinguishable. The puzzle, said Chomsky, is that children never make the linear error. In 1987 Stephen Crain and Mineharu Nakayama tested children aged three to five by asking them to form questions from sentences like this and found no linear errors. Critics reply in three ways. First, the premise: Geoffrey Pullum and Barbara Scholz (2002) argued that such evidence is more common than claimed, and that "learning from the data" is not hopeless. Second, statistical learners: Florencia Reali and Morten Christiansen (2005) showed simple models trained on child-directed speech get most of these questions right with no built-in syntax. Third, the whole picture: Michael Tomasello and the usage-based school argue that children begin with specific phrases ("Where's the ___?", "I wanna ___") and gradually generalise them into abstract patterns, using general social and pattern-finding skills (reading others' intentions, categorising). The nativists reply that those models learn from far more or differently structured data than a child gets, and that cross-linguistic universals need explaining. The debate is live and has moved. Chomsky's own proposals have shrunk the built-in part to little more than a combining operation (Merge), and Tomasello's side now accepts a rich set of innate learning skills. Language models, trained on enormously more text than any child hears, are now a new kind of evidence; what they show about children is argued by both camps (see Computational Linguistics, 36/30). Daniel Everett's claim that Pirahã in Brazil lacks recursion (2005) was meant to challenge the universal-grammar picture; Andrew Nevins and colleagues (2009) disputed his data. A fair summary: nobody now thinks the child is a blank slate, and nobody agrees about how much of grammar is built in. Critical periods: Genie, Nicaraguan Sign Language and the age curve Is there a window for language learning? The biologist Eric Lenneberg argued in 1967 that language is easily acquired in childhood and, after about puberty, no longer in the same way. Animal studies of imprinting and song learning in birds gave the idea a model. Humans can obviously not be experimented on, so we use accidents of history. The best known is Genie. She was born in 1957 in Los Angeles and from about the age of 20 months was kept confined by her father, mostly alone, with almost no speech around her. She was discovered in November 1970, aged 13 years and 7 months, and linguists including Susan Curtiss studied her for years. Genie learned many words and some structure, but her grammar stayed limited and she never fully acquired a first language. It is tempting to treat this as proof of a critical period. It is not clean evidence: she had suffered severe abuse, malnutrition and trauma, and it is impossible to separate a missing window from everything else; Curtiss argued against a general intellectual impairment, others were less sure. The research itself has been criticised on ethical grounds, since the people studying her were also her carers. The testing ended in 1978, and her later years were hard. Genie is a case to tell with care and without sensationalism. Better evidence comes from people who come to a language late, but without being deprived. Jacqueline Johnson and Elissa Newport (1989) tested Chinese and Korean immigrants to the United States on English grammar: those who arrived before about seven years old matched native speakers, performance fell as the age of arrival rose through adolescence, and after puberty it was scattered. In 2018 Joshua Hartshorne, Joshua Tenenbaum and Steven Pinker analysed an online English grammar quiz taken by roughly two-thirds of a million people. Their results fit a critical period that ends much later than expected: the ability to learn grammar stays high until about 17 or 18, then declines. They suggested cultural and social changes at that age might be partly responsible, in addition to brain maturation. Rachel Mayberry and colleagues found that deaf people who met their first language late (a sign language, for deaf children of hearing parents with limited access) showed lasting effects on grammar, even compared with those who learned a second language later. The effects are about first language exposure. Interactive · Three shapes the age curve could have Age of first exposure to the language Sharp cut-off at 12   Steady decline   High until 17, then falling Schematic shapes, not data. They show how different the three predictions are for, say, a 14-year-old. Real studies measure noisy scores, and the end age and steepness depend on what is measured (accent, grammar, vocabulary). Without JavaScript: a sharp cut-off predicts high scores until the cut-off age, then low; a steady decline predicts a score falling gradually from birth; the 2018 results resemble a plateau into the late teens followed by a decline. The most striking evidence is not a deprivation but a creation. Before 1977 deaf Nicaraguan children mostly lived at home, each with their own family's gestures. When special schools opened in Managua, starting in 1977 with about 50 deaf children and growing to more than 400 pupils in two schools by 1983, the children were together for the first time. They did not receive a sign language from teachers (who used Spanish and fingerspelling). They made one. Judy Kegl, invited by the ministry of education in 1986, and later Ann Senghas and Marie Coppola, documented how each cohort of younger children made the language more systematic. Senghas, Kita and Özyürek (2004) found that younger signers split a rolling-down-a-hill motion into separate signs for manner and path, while the older cohort's signing tended to express both at once, as hearing people's gestures do. A similar story is told for Al-Sayyid Bedouin Sign Language in the Negev, which arose within about three generations (Sandler and colleagues, 2005). What it shows is debated: nativists see children imposing structure; others stress that a community, its needs and its transmission across cohorts matter. Bilingual children: myths and evidence Perhaps half the world's people use more than one language in their daily lives (a common estimate by the psycholinguist François Grosjean). Bilingualism is the normal case, not a puzzle, and the research on bilingual children has gone through a cycle of myth, over-enthusiasm and correction. The old worry, that two languages confuse a child and delay speech, is not supported: bilingual children reach the main milestones in a similar window. What the evidence does show is that vocabulary depends on input. A bilingual toddler usually knows fewer words in each language than a monolingual of the same age, because she hears each for part of the day, but her total across both languages is comparable (Hoff and colleagues, 2012). Mixing languages in a sentence (code-switching) is not a sign of confusion: Shana Poplack's 1980 study of Spanish–English speakers in New York found it follows grammatical constraints and is commonest among the most fluent speakers. The more recent "cognitive advantage" claim, that bilinguals have a stronger executive control, went from headline to dispute: Minna Lehtonen and colleagues (2018) pooled 152 adult studies and found no advantage after correcting for publication bias. Similar disagreement attends the finding that bilinguals show Alzheimer's symptoms about four years later (Bialystok and colleagues, 2007), since later studies with better controls differ. Try your hand at sorting the claims. Interactive · Myth or evidence? Your score is kept in the page only while it is open. Verdicts summarise the studies named in the text. Without JavaScript: bilingualism confusing children is not supported; smaller vocabulary per language but similar totals is supported; code-switching as confusion is not supported; an executive-function advantage and delayed dementia are mixed; most of the world using several languages is supported. Practical points the research does support: amount and quality of input matter more than any one-parent-one-language rule; bilingualism does not cause a language disorder, and a bilingual child with a delay should be assessed in both languages by someone who knows the field. The evidence is about averages; any particular family has its own situation. Sign languages are full languages Until the 1960s many hearing people, including many linguists, treated sign as gesture or pantomime or a broken English on the hands. In 1960 William Stokoe, a professor of English at Gallaudet, published Sign Language Structure, showing that American Sign Language (ASL) signs are built from smaller meaningless units, as spoken words are built from sounds: where the hands are (location), their shape (handshape), the way they move (movement), and later also palm orientation. Change one, and you change the word. Robbin Battison's 1974 minimal pairs are standard examples. Try them. Interactive · Change one part, change the word Pairs follow standard textbook descriptions (Battison 1974, as summarised in references on ASL phonology). Signs are described in words here; a Deaf-led signing course or the free ASL resources linked below are the way to see them. Without JavaScript: APPLE and ONION differ in location; MOTHER and FATHER differ in location (chin or forehead); CAR and WHICH in handshape; PROPOSE and MAYBE in movement; TEMPORARY and TRAIN in palm orientation. Everything else follows. Sign languages have grammar, new vocabulary, poetry, jokes and dialects; they use space and the face for grammar (raised eyebrows can mark a yes/no question) and can express two things at once, which speech cannot. Deaf children of Deaf parents learn them on roughly the spoken timetable, including "babbling" with their hands (Laura-Ann Petitto and Paula Marentette, 1991). Damage to the left hemisphere impairs signing as it impairs speech, and signers with strokes can lose grammar while keeping the ability to gesture (Ursula Bellugi, Howard Poizner and Edward Klima's work on sign aphasia). Sign languages are many and not universal: ASL descends from French Sign Language and is unrelated to British Sign Language, and the two are not mutually intelligible. Counts of sign languages vary widely (from well over a hundred to over three hundred) because it depends on what counts as separate. Most deaf children, about nine in ten by a commonly cited estimate, are born to hearing parents who have never met a sign language. That is why the age-of-exposure research is not abstract: it concerns real children who sometimes reach school with no full language. How best to provide early access to language, through sign, spoken language or both, is discussed in families and Deaf communities, and this piece does not advise on it. What a graduate just sees Babies are statistical learners before they know any word, and "break the stream where predictability dips" is a real strategy. "Goed" is a sign that a rule has been learned; most irregular verbs are produced correctly most of the time. The poverty-of-the-stimulus argument is about what the data can and cannot tell a learner; both camps now accept some innate machinery, and disagree about how much and of what kind. Genie is a hard case with many confounds, not a clean test; the best evidence for a sensitive period comes from late-exposed learners. Deaf children made a language in Nicaragua; languages are made by communities of children. Bilingualism is normal, does not delay children, and its "cognitive benefits" are disputed. Sign languages are languages, with phonology, grammar and a brain. Where it shows up next Psycholinguistics & the Brain (36/28): how an adult brain handles the language the child built. Computational Linguistics (36/30): what language models learn, and how it compares to what children learn from far less. Sociolinguistics (36/26): language shift, heritage speakers and the politics of which languages children are taught in. In the world: early-years teaching, speech-language therapy, deaf education, language policy in schools, and the design of reading programmes. If you want to go deeper 📘 Textbook. Eve Clark, First Language Acquisition (Cambridge University Press, 3rd ed. 2016) is a standard course text; Barbara Lust, Child Language: Acquisition and Growth (Cambridge) is another. 📘 Gentler. Steven Pinker, The Language Instinct (1994) argues the nativist case in an engaging way; Michael Tomasello, Constructing a Language (2003) is the usage-based reply; Susan Curtiss's account of Genie and Russ Rymer's Genie: A Scientific Tragedy (1993) give the case in full. 🌐 Free text. The Noba Project and OpenStax psychology texts have free modules on language development; CHILDES (childes.talkbank.org) is a free archive of recorded child speech. 🌐 Free course. MIT OpenCourseWare 24.900 Introduction to Linguistics (Spring 2022) includes sessions on language acquisition and on language change. The takeaway Children build a language rather than copy one: they find words by tracking statistics, invent rules and over-apply them, and in some conditions create new languages. How much of that rests on innate structure and how much on general learning is the field's great open argument. The age window is real but the best evidence for it is more careful, and later, than the Genie story suggests. Try this: next time you hear a small child say a "wrong" past tense or plural, write it down. Then check whether it follows a rule they have likely learned. Sources & Further Reading Sources. Saffran, Aslin and Newport, Science 1996; Hauser, Newport and Aslin, Cognition 2001. Berko, Word 1958; Marcus et al., Monographs of the SRCD 1992 (11,521 uses; median 2.5%); Maslen et al. 2004 (4.2% mean in a dense corpus); Maratsos 2000; Rumelhart and McClelland 1986; Pinker and Prince 1988; Slobin on Turkish. Chomsky 1959, 1980; Crain and Nakayama, Language 1987; Pullum and Scholz 2002; Reali and Christiansen 2005; Tomasello 2003; Everett 2005; Nevins, Pesetsky and Rodrigues 2009. Lenneberg 1967; Genie: Curtiss 1977, Rymer 1993, Wikipedia "Genie (feral child)" (born 1957, found November 1970 aged 13 years 7 months, confined from about 20 months, testing ended 1978); Johnson and Newport, Cognitive Psychology 1989; Hartshorne, Tenenbaum and Pinker, Cognition 2018; Mayberry and Lock 2003. Nicaraguan Sign Language: Wikipedia "Nicaraguan Sign Language" (1977 school with about 50 children, more than 400 pupils in 1983, Kegl 1986); Senghas, Kita and Özyürek, Science 2004; Sandler et al., PNAS 2005. Grosjean on bilingualism; Hoff et al., Journal of Child Language 2012; Poplack 1980; Lehtonen et al., Psychological Bulletin 2018 (152 studies); Bialystok, Craik and Freedman, Neuropsychologia 2007. Stokoe 1960; Battison 1974; Petitto and Marentette, Science 1991; Poizner, Klima and Bellugi 1987. The numbers the toys display were computed in node; they are illustrative. This piece describes research on children and is not clinical or educational advice. 📘 Clark, First Language Acquisition 📘 Pinker, The Language Instinct 🌐 CHILDES Step 11. Psycholinguistics & the Brain: How a Mind Handles a Sentence --------------------------------------------------------------------- Why: How the brain produces and understands language. In under half a second you recognise a word, guess the next one and sometimes get led up the garden path. The classic brain map of it is now being redrawn. Someone says, "She spread the warm bread with…" and you already know the next word, before it arrives. When they say "butter", nothing happens. When they say "socks", something does: about 400 milliseconds after the word begins, your brain gives off a small electrical signal that a scalp electrode can pick up. Nobody told your brain to do this. It was running a few words ahead of the speaker and found itself wrong. Psycholinguistics is the study of what the mind does in real time with language: recognising a word among tens of thousands in a few hundred milliseconds, building a sentence structure word by word, reading a writing system that has existed for a tiny fraction of human history, and, at the edges, asking whether the language you speak shapes how you think. Neurolinguistics asks where in the brain it happens. This piece is descriptive. It explains research on how language works in healthy brains and after injury, and it is not medical advice. The course in one box Who takes itLinguistics and psychology majors in year 3 or 4; also speech-language pathology, neuroscience, education and computer-science students. It is a standard course in programmes that follow the LSA and UK subject-benchmark guidance, and in cognitive-science majors. What it is forLinking the structures linguists describe to the processes and brain regions that produce and understand them. Comes afterSyntax (36/23) and Cognitive Psychology (35/22). Leads toComputational Linguistics (36/30), where models of processing become programs, and Clinical Psychology (35/27) and Biological Psychology (35/20), which it sits beside. What this piece gives youYou will not be able to read an EEG or assess a patient. You will know the classic brain map and why it is being redrawn, how word recognition and sentence parsing work in time, what dyslexia is and is not, and what the language-and-thought evidence shows. Easier first reads: Brain imaging: what a scan can and cannot show inside Biological Psychology (35/20) and Language and thought in Cognitive Psychology (35/22). Not medical advice. This explains research on language and the brain. If you or someone you know has had a stroke or injury, or has trouble reading or speaking, a doctor or speech-language pathologist is the person to ask. In an emergency call your local emergency number. Broca, Wernicke and what imaging changed In 1861 the French physician Paul Broca met a patient, Louis Victor Leborgne, known as "Tan" because that was nearly the only syllable he could say, although he seemed to understand speech. After Leborgne died Broca found damage in the left frontal lobe. In 1874 the German neurologist Carl Wernicke described patients with the opposite pattern: fluent, grammatical-sounding speech that made little sense, and poor understanding, after damage farther back in the left temporal lobe. Put together, they gave the classic picture, later developed by Lichtheim (1885) and Norman Geschwind (1960s): a front area for producing speech, a back area for understanding it, and a fibre bundle (the arcuate fasciculus) joining them. That picture still organises how clinicians describe the aphasias, language disorders after brain damage. Three questions sort most patients: is speech fluent, is comprehension good, can they repeat what they hear? Here is the standard (Boston) classification as a decision table. Interactive · Sort a pattern into a syndrome SpeechFluentEffortful, halting UnderstandingGoodPoor Repeating what is saidGoodPoor Eight combinations, eight classic labels. Real patients rarely fit one box exactly, and lesion sites here are "typical", not guaranteed. Without JavaScript: fluent, poor comprehension, poor repetition is Wernicke’s; effortful, good comprehension, poor repetition is Broca’s; fluent, good comprehension, poor repetition is conduction aphasia; effortful and poor on both is global; fluent with good understanding and repetition is anomic aphasia. Modern imaging made the map more complicated. In 2007 Nina Dronkers and colleagues scanned the preserved brain of Leborgne with high-resolution MRI. The damage reached well beyond Broca's area into deeper structures and white-matter tracts, which fits a later finding that damage to the underlying connections may predict lasting speech problems better than damage to the cortical area alone. Functional MRI and diffusion imaging show language using a network, not two spots: regions in left frontal, temporal and parietal cortex, linked by several pathways. Gregory Hickok and David Poeppel's dual-stream model (2007) proposes a ventral route from sound to meaning and a dorsal route from sound to articulation. Evelina Fedorenko and colleagues find that, in each person, a set of regions responds strongly to language and much less to maths, music or general effort. Aphasia and the language network also look separate from thought: Rosemary Varley and colleagues (2005) showed that people with severe grammar loss could still solve algebra problems. And for most people, the left hemisphere is dominant for language, but not for everyone; a minority, more often left-handers, have language on the right or both sides. Broca and Wernicke were right that language has a biological address, and the address has turned out to be a neighbourhood. Hearing a word: a race among competitors An adult knows a lot of words. Brysbaert and colleagues (2016) estimated that a typical 20-year-old American English speaker knows about 42,000 lemmas (dictionary forms), and a speaker hears several words every second. Yet word recognition is very quick, often complete within about a quarter of a second of the word starting, sometimes before the speaker finishes. How can the mind pick one word out of so many that fast? The answer in cohort models (William Marslen-Wilson, 1987) is that it does not wait. As soon as the first sound arrives, every word beginning that way is activated, the cohort. Each new sound knocks out the words that no longer fit. The point at which only one word remains is the uniqueness point, and recognition tends to happen close to it. The TRACE model (McClelland and Elman, 1986) adds competition: active words inhibit each other. Frequent words win faster, and words with many similar-sounding neighbours (cat, cap, cab…) are slower to recognise than isolated ones. Step through a word below. Interactive · Narrow the cohort Word being spoken Sounds heard so far A toy lexicon of 26 American English words, written with ARPAbet-style sound symbols (K as in "cat", AE as in "cat", AH as in "but"). A real lexicon has tens of thousands of words. The effect of frequency and neighbours is described in the text, not simulated. Without JavaScript: in the toy lexicon all 26 words begin with the sound K; after K AE thirteen remain; after K AE P four (cap, captain, capital, captive); captain becomes unique only at its fifth sound, when AH separates it from captive. Cap and cat never become unique by sound alone because longer words begin with them. Notice what the toy shows. Cap is a prefix of other words, so it is never "unique" at its end by sound alone. Listeners use the pause after it, stress, and the sentence. Context is not a luxury in recognition; it is part of how the race ends. Brain-imaging studies place much of this in the left temporal cortex, with the signature of "word found" appearing quickly after the point of recognition. Garden paths: the parser commits early Read: The horse raced past the barn fell. Most readers stumble at "fell". Nothing is ungrammatical. The sentence means "the horse (that was) raced past the barn fell", but your mind took "raced" as the main verb of the sentence as soon as it appeared, and had to rebuild when it met a second verb. Thomas Bever described such garden-path sentences in 1970. Others: While Anna dressed the baby slept. (people commit to "Anna dressed the baby", and Christianson and colleagues, 2001, found that readers often keep the wrong reading even after they reread: asked "Did Anna dress the baby?", many say yes.) And The old man the boats. (the old, as a noun, man the boats.) These sentences matter because they show that sentences are not read all at once, and that the parser is incremental: it builds a structure word by word and commits. Two families of theory disagree on how it chooses. The garden-path model (Lyn Frazier and Keith Rayner, 1982) says that a first-pass process picks the simplest structure, using only syntactic categories; meaning corrects it afterwards. Constraint-based accounts (Maryellen MacDonald, Neal Pearlmutter and Mark Seidenberg, 1994; John Trueswell and Michael Tanenhaus) say that many sources of information, such as verb preferences, meaning and context, compete from the start. Evidence for the second includes eye-tracking: "The defendant examined by the lawyer" tricks readers, but "The evidence examined by the lawyer" doesn't, since evidence cannot examine anything (Trueswell, Tanenhaus and Garnsey, 1994). And in the visual-world method (Tanenhaus and colleagues, 1995), listeners' eyes dart to objects as they are named, which shows parsing in action. Also, Fernanda Ferreira has argued comprehension is often "good enough": we settle for a plausible meaning, not a fully worked-out structure. Try the effect on yourself. Interactive · Read word by word and time yourself StartNext word Tap Next word as soon as you have read each word. Do the tricky version and its easier twin, then compare the times at the highlighted word. One person on one screen is a demonstration, not a study: the real effect is found by averaging many readers and many sentences, and tap speed and attention add a lot of noise. Nothing is stored or sent. Without JavaScript: in "The horse raced past the barn fell" readers slow down at "fell"; the version "The horse that was raced past the barn fell" removes the trap. Likewise a comma in "While Anna dressed, the baby slept" removes the garden path. Reading is not natural: two routes and dyslexia Speaking is learned without lessons. Reading is not. Alphabets are a few thousand years old, and most human languages have never been written. Reading has to be built on brain systems that evolved for something else. Stanislas Dehaene's "neuronal recycling" idea is that a patch of left visual cortex (the visual word form area, identified by Laurent Cohen and colleagues in 2000) is recruited for recognising letter strings; Dehaene and colleagues (2010) found that adults who learn to read late show changes in visual and language regions. How do we get from print to sound? The dual-route model (Max Coltheart and colleagues) proposes two pathways. A lexical route looks up whole words in memory, which is the only way to read irregular words like pint or have. A sub-lexical route converts letters into sounds by rules, which is the only way to read a made-up word like frib. Competing "triangle" models (Seidenberg and McClelland, 1989; Plaut and colleagues, 1996) use one network trained on spelling, sound and meaning. Brain damage in adults has produced patterns that fit the two-route picture: surface dyslexia (irregular words regularised, as if pint rhymed with mint) and phonological dyslexia (real words fine, invented words impossible). Switch routes off in the tool and see. Interactive · Which route reads this word? Lexical route on Sub-lexical route on The rule route is a small letter-to-sound program (one vowel, optional silent e) run live; the "real" column is standard American English. The routes are model components: damage in real people is partial and messy. This illustrates acquired dyslexias, which differ from developmental dyslexia. Without JavaScript: with only the rule route, pint is read like "mint", have like "gave" and love like "cove"; with only the lookup route, invented words such as frib cannot be read. Developmental dyslexia is a difficulty with accurate and fluent word reading and spelling that persists despite normal instruction and ordinary intelligence. Estimates of how common it is vary with the definition, commonly about 5 to 10 percent of people. Its most replicated core is a weakness in phonological processing, seeing the sounds inside words (the phonological-deficit account, associated with Margaret Snowling, among others), and it runs in families. Some myths are worth retiring: dyslexic people do not see letters backwards (young children of all kinds reverse letters), it is not a sign of low intelligence, and coloured lenses and vision therapy are not supported as treatments (a joint statement of the American Academy of Pediatrics and ophthalmology bodies, 2009, reaffirmed 2011). What does help is explicit, structured teaching of the link between sounds and letters, started early. Writing systems matter too: Philip Seymour and colleagues (2003) found that children learning regular alphabets such as Finnish, Italian and German learn to decode accurately within about a year, while English, with its irregular spelling (a legacy of the sound changes in Historical Linguistics, 36/25), takes much longer; in regular systems dyslexia shows up more as slow reading than as inaccurate reading. Chinese, which writes meaning-bearing characters, taxes different brain systems. Whether to seek an assessment, and what support is available, depends on the country and school; that is a question for a qualified professional. Prediction: the brain runs ahead of the words Back to the bread and the socks. In 1980 Marta Kutas and Steven Hillyard recorded electrical activity from people's scalps as they read sentences one word at a time and found that a semantically odd final word ("He spread the warm bread with socks") evoked a negative wave peaking about 400 milliseconds after the word began, the N400. The wave is graded: the less expected a word, the bigger it is, and it appears even for words that are perfectly sensible but unlikely. Researchers use cloze probability, the share of people who would fill in a given word, to measure expectation, and information theory to describe it: the surprisal of a word is −log₂ of its probability. Lower probability means more bits of surprise. Studies (including Stefan Frank and colleagues, 2015) find N400 size tracks surprisal, though what the wave reflects (difficulty of integrating a word, ease of retrieving it, or both) is debated by Kutas, Federmeier and others. Interactive · How surprised is the brain? Cloze probabilities are invented for illustration (a probability of 0.0005 stands for "almost nobody"). The waveform is a schematic: depth is proportional to surprisal, as a modelling assumption, and it is not a recording. Negative is drawn upward, as in much of the ERP literature. Without JavaScript: "butter" after "spread the warm bread with" is expected (about 0.3 bits of surprisal); "jam" is less so (about 3 bits); "socks" is almost never said (about 11 bits), so the N400 is large. Prediction is not an optional extra. It explains why a noisy phone line is still understandable, why you can finish a friend's sentence, and why a garden path hurts: the parser bet and lost. It is also the training task of today's language models, which learn by predicting the next word. That similarity is real; whether it means the models and brains do the same thing is argued (see Computational Linguistics, 36/30). Does language shape thought? Weak, strong and the evidence Benjamin Lee Whorf, writing in the 1930s and 1940s, and his teacher Edward Sapir, are credited with the idea of linguistic relativity: the language you speak affects how you think. The strong version, that language determines thought, so people cannot think what they cannot say, is rejected by nearly everyone. Whorf's own best-known case, that the Hopi had no concept of time, was challenged by Ekkehart Malotki's 1983 book, which documented Hopi time expressions. The weak version, that language nudges attention, memory and speed of categorisation, has real evidence. Three areas of that evidence are worth knowing. Colour. Berlin and Kay (1969) argued that colour terms across languages grow in a near-fixed order: dark and light first, then red, then green or yellow, then blue, then brown, then purple, pink, orange and grey. That suggests biology constrains the vocabulary (the later World Color Survey, covering 110 unwritten languages, found the pattern as a strong tendency, not a law). But the vocabulary then affects perception a little. Winawer and colleagues (2007) found that Russian speakers, whose language has separate basic words for light blue (goluboy) and dark blue (siniy), were faster at telling apart two blues that fall on opposite sides of that line, and the advantage vanished when they had to hold a string of digits in mind at the same time, which suggests the effect works through language in the moment. Roberson and colleagues (2005) found a similar boundary effect in speakers of Himba, whose language does not split blue from green. The effects are modest, measured in milliseconds, and some replications are mixed. Interactive · Berlin and Kay's colour ladder Basic colour terms in the language The order is Berlin and Kay’s (1969), as revised by Kay and Maffi (1999) and the World Color Survey. At stage 3-4 and 5 either green or yellow can come first, and some languages have more than 11 (Russian has 12). The swatches are only approximate screen colours. Without JavaScript: two terms (dark and light), three adds red, four adds green or yellow, five both, six adds blue, seven brown, and eight to eleven purple, pink, orange and grey. Space. Speakers of Guugu Yimithirr, in northern Queensland, do not say "left" or "right". They give compass directions for everything, even small things nearby: "there is an ant on your southwest leg". Stephen Levinson and colleagues showed that speakers of such languages keep track of compass directions even in unfamiliar surroundings, and that people laying out picture cards for time put them in order from east to west whatever way they face (Boroditsky and Gaby, 2010, for Kuuk Thaayorre). Critics, such as Li and Gleitman (2002), found that English speakers' choices flip if the environment gives cues, and argued that the language is not the cause; Levinson's group replied. The jury is still out, but the difference in everyday description is real. Interactive · Left of me, or east? Facing northFacing eastFacing southFacing west Without JavaScript: the cup is always due east. A speaker who uses "left" and "right" must change the description every time they turn; a speaker of a compass-direction language does not. Number. The Pirahã of the Amazon and the Mundurukú of Brazil have few or no exact number words. Peter Gordon (2004) and Pierre Pica and colleagues (2004) found that their speakers estimate small quantities well but become inaccurate for exact larger ones, which suggests that number words are cultural tools that help exactness. Claims about grammatical gender affecting how people describe objects have been harder to replicate. A fair summary: language does not set the limits of thought, but it can shift attention, speed and memory in specific tasks, and each claim has to be tested one by one. What a graduate just sees Language lives in a left-hemisphere network, not in two spots; Broca and Wernicke named regions of a larger system. Aphasia types can be sorted by three questions: fluency, comprehension, repetition. You recognise a word before it ends because you track a shrinking cohort of candidates. Parsing is incremental: the mind commits early, and a garden path is the cost of that bet. Reading is a recycled skill; irregular words need lookup, new words need rules, and dyslexia is a phonological difficulty, not a vision problem. The brain predicts the next word constantly; an unexpected one costs about 400 ms of extra effort. Language nudges thought in specific, modest ways; it does not cage it. Where it shows up next Computational Linguistics (36/30): surprisal, parsers and next-word models are the same ideas as programs. Clinical Psychology (35/27) and Biological Psychology (35/20): aphasia, stroke recovery and the limits of brain scans. Language Acquisition (36/27) and Sociolinguistics (36/26): how the system gets built, and how it varies between groups. In the world: speech-language therapy, reading instruction, accessible text design, brain-computer interfaces for speech, and the design of user interfaces that avoid garden paths. If you want to go deeper 📘 Textbook. Trevor Harley, The Psychology of Language: From Data to Theory (Psychology Press / Routledge, 4th ed. 2014); Matthew Traxler, Introduction to Psycholinguistics (Wiley-Blackwell, 2nd ed. 2022). 📘 Gentler. Stanislas Dehaene, Reading in the Brain (2009); Guy Deutscher, Through the Language Glass (2010) on language and thought; Maryanne Wolf, Proust and the Squid (2007) on the reading brain. 🌐 Free text. The National Aphasia Association (aphasia.org) and the British Dyslexia Association (bdadyslexia.org.uk) publish plain-language guides; the Noba Project has free modules on language. 🌐 Free course. MIT OpenCourseWare 24.900 Introduction to Linguistics (Spring 2022) covers the structure of language this subject builds on; for the brain side, MIT OpenCourseWare also hosts courses in Brain and Cognitive Sciences (course 9). The takeaway Understanding language is a race against time: a word is found among thousands in a quarter of a second, a sentence is built before it ends, and the next word is guessed before it arrives. The old two-area brain map was a good first draft of a network. Language shapes thought modestly and specifically, and reading is a hard-won skill that depends on sound. Try this: when someone is mid-sentence, try to guess their next word, and notice how often you are right. Or find a garden-path sentence in a headline (they are common) and notice where you stumbled. Sources & Further Reading Sources. Broca 1861; Wernicke 1874; Lichtheim 1885; Geschwind 1965; Dronkers et al., Brain 2007 (Leborgne); Hickok and Poeppel, Nature Reviews Neuroscience 2007; Fedorenko et al. 2010; Varley et al., PNAS 2005; Boston classification as in standard neurology texts (Goodglass and Kaplan). Brysbaert et al., Frontiers in Psychology 2016 (about 42,000 lemmas); Marslen-Wilson 1987; McClelland and Elman 1986. Bever 1970; Frazier and Rayner 1982; Christianson et al., Cognitive Psychology 2001; MacDonald, Pearlmutter and Seidenberg 1994; Trueswell, Tanenhaus and Garnsey 1994; Tanenhaus et al., Science 1995; Ferreira 2003. Coltheart et al. 2001; Seidenberg and McClelland 1989; Plaut et al. 1996; Cohen et al. 2000; Dehaene et al., Science 2010; Seymour, Aro and Erskine, British Journal of Psychology 2003; Snowling on the phonological-deficit account; the 2009 joint statement on dyslexia, vision and learning disabilities (American Academy of Pediatrics with the American Academy of Ophthalmology and others), reaffirmed 2011. Kutas and Hillyard, Science 1980; Kutas and Federmeier 2011; Frank et al. 2015. Whorf 1940; Malotki 1983; Berlin and Kay 1969; Kay and Maffi 1999; Kay et al. 2009 (World Color Survey, 110 languages); Winawer et al., PNAS 2007; Roberson et al. 2005; Levinson 1997 and 2003; Li and Gleitman 2002; Boroditsky and Gaby, Psychological Science 2010; Gordon, Science 2004; Pica et al., Science 2004. Cloze and surprisal numbers in the toy are invented; its arithmetic was run in node. This piece describes research and is not medical or educational advice. 📘 Harley, The Psychology of Language 📘 Dehaene, Reading in the Brain 🌐 National Aphasia Association Step 12. Computational Linguistics: From Grammars to Language Models -------------------------------------------------------------------- Why: From grammars to language models. Teaching machines to handle language began with hand-written rules, moved to counting, and now rests on neural networks trained on a large share of the written web. Each stage learned something about language. What the latest stage proves is still argued. Ask a phone to translate "the trophy would not fit in the suitcase because it was too big" into French or Spanish, where the pronoun must be masculine or feminine to match one of the two objects, and the system must decide in a fraction of a second what a human does without noticing. Change one word at the end ("too small") and the answer flips. Nothing in the spelling tells you; the knowledge is about trophies and suitcases. Computational linguistics sits where that problem lives. It asks how to represent words, structure and meaning in forms a program can use, and how to tell whether the program got it right. Its practical side, usually called natural language processing, produced search, spelling correction, translation, speech recognition and, lately, chatbots. Its scientific side asks what these systems reveal about language itself. A graduate keeps both questions in view. The course in one box Who takes itLinguistics majors, usually in the final year as the bridge to language technology, and computer science students who choose a language track (courses such as Stanford's CS224N sit in computer science, with slides free online). It appears in language-technology master's degrees in many countries. What it is forRepresenting language so a computer can analyse or produce it, and judging how well it does. Comes afterSyntax (36/23), for trees and ambiguity, and Probability & Statistics (34/05), for counts, probabilities and error bars. Leads toIt sits beside Psycholinguistics & the Brain (36/28), Semantics & Pragmatics (36/24) and Machine Learning (34/32). What this piece gives youYou will not be able to train a model. You will see why "word" is not a simple unit, why counts already give you a crude language model, how a parser deals with ambiguity, what an embedding is, what attention looks up, why a metric is only a proxy, and which claims about language models are firm and which are not. The easier first read is How Large Language Models Actually Work (29/02); this piece goes sideways into the linguistics. All texts here are public domain or invented. Tokenisation: what counts as a word? Before any model can read, text must be cut into units. English spelling makes it look easy: split at spaces. But "don't" is two words in one, "New York" is one in two, and many languages never put spaces in the places English does. Mandarin writes no spaces between words. Turkish and Swahili pack a whole English sentence's worth of structure into one written word. German glues nouns together. LanguageWritten word, split into morphemesGlossFree translation Turkishev-ler-iniz-denhouse-PL-2PL.POSS-ABLfrom your (plural) houses Swahilini-na-ku-pend-a1SG-PRS-2SG.OBJ-love-FVI love you Mandarinwǒ xǐhuan kāfēi (我喜欢咖啡)1SG like coffeeI like coffee Leipzig-style glosses: PL plural, POSS possessive, ABL ablative (“from”), SG singular, PRS present, OBJ object, FV final vowel. Modern systems avoid choosing words by hand. Byte-pair encoding (BPE), borrowed from data compression (Philip Gage, 1994) and brought to translation by Sennrich, Haddow and Birch (2016), starts from single characters and repeatedly merges the most frequent adjacent pair into a new unit. Common words become single tokens; rare words are spelled out from pieces. Tokenisers for large models typically have tens of thousands to a few hundred thousand pieces (GPT-2 had 50,257). The try-it below runs the real algorithm on a tiny invented corpus. Interactive · Learn a vocabulary by merging Merge onceMerge 5Run to the endReset Invented corpus of 30 words: walking ×4, walked ×3, walks ×3, talking ×4, talked ×3, talks ×2, talk ×5, walk ×6. The vocabulary is the starting symbols plus one new piece per merge. Ties are broken alphabetically. Without JavaScript: the corpus starts as 191 character tokens from 12 symbols. After 10 merges the vocabulary is 22 and the corpus 59 tokens, with walk, talk and the endings -ing, -ed having emerged as units; after 18 merges every word is a single token (30 tokens). Notice the trade: a bigger vocabulary shortens every text but leaves rare pieces under-trained. And tokenisers learn mostly from the languages in their training text, so the same sentence can take several times as many tokens in some languages as in English (Petrov et al., 2023; Ahia et al., 2023), which affects cost and quality for those speakers. N-grams and Zipf: counting your way to a language model The oldest statistical language model just counts. An n-gram model predicts the next word from the previous n − 1 words by looking at how often each continuation followed them in a corpus. Claude Shannon showed in 1948 how sequences generated this way start to look like English. The model is crude, but it is a real language model, and its mathematics (probabilities of sequences, perplexity, smoothing) carried into neural models. There is a catch hiding in any word count, first described by the linguist George Kingsley Zipf (Human Behavior and the Principle of Least Effort, 1949). A few words are extremely common and most are rare, and the pattern is regular: the second-commonest word is about half as frequent as the commonest, the third about a third, and so on. Plot rank against frequency on log axes and you get a roughly straight line of slope near −1. The consequence for language modelling is harsh: however much text you collect, many word sequences will have never appeared, and the model must still give them a probability. The demo below runs on the first 1,235 words of Lewis Carroll's Alice's Adventures in Wonderland (1865, public domain, from Project Gutenberg). Interactive · Zipf, and a model that only counts What comes next? Write with it BigramTrigram Generate again The passage is tiny, so the counts are noisy and the tail of single words is long. The generated text starts at a random place and picks each next word with the probability seen in the passage; the trigram model mostly copies the original, because with this little text most two-word contexts have only one continuation. Without JavaScript: in the passage “the” is 4.5% of all words, then “to”, “she”, “was” and “and”; the fitted log-log slope is about −0.8; 246 of the 420 different words occur exactly once. After “the”, the commonest next word is “rabbit”. In 1957 Noam Chomsky used “Colorless green ideas sleep furiously” to argue that grammaticality cannot be reduced to probability, since the sentence is grammatical yet has presumably never occurred. Fernando Pereira (2000) answered that a statistical model with word classes does rate it far likelier than the same words reversed, so probability and grammar are not as opposed as the argument assumed. Both sides learned something: the models needed classes and smoothing, and the grammarians needed to say what a probability is a probability of. Parsing: finding the structure, and its ambiguity A parser assigns a structure to a sentence. In constituency parsing the structure is a tree of phrases; in dependency parsing it is a set of word-to-word links. The big shift of the 1990s was to learn grammars from annotated text: the Penn Treebank (Marcus, Santorini and Marcinkiewicz, 1993) gave thousands of hand-parsed sentences, and today the Universal Dependencies project (release 2.18, May 2026) offers over 200 treebanks in over 150 languages with one shared annotation scheme, which is how a parser trained on one language can be compared with another. The core problem is ambiguity. The grammar of English lets one string have several trees, and a parser must choose. A probabilistic context-free grammar (PCFG) attaches a probability to each rule and scores a tree as the product of its rules. The demo uses a tiny hand-made grammar and a real chart parser that finds every tree. Interactive · Which tree does the grammar prefer? She saw the man with the telescopeOld men and women left How strongly this grammar likes attaching “with…” to the verb Invented grammar: each noun-phrase rule and the verb rules have fixed probabilities except the slider, which moves probability between “verb + noun phrase” and “verb phrase + prepositional phrase”. A real parser would learn these numbers from a treebank. The slider has no effect on the second sentence. Without JavaScript: “she saw the man with the telescope” has two parses. With a 10% preference for verb attachment, the telescope belongs to the man; at 30% the telescope belongs to the seeing. The break-even is 15%. “Old men and women left” has two parses of equal probability: the grammar cannot choose, and a human would need context. The second sentence shows the limit of structure alone. Choosing between “old (men and women)” and “(old men) and women” takes knowledge about the world, which is why syntax-only parsers hit a ceiling. See Syntax (36/23) for the linguistic side of ambiguity. Word vectors: meaning from company J. R. Firth wrote in 1957, “You shall know a word by the company it keeps.” Zellig Harris's distributional hypothesis of the 1950s is the same idea: words that occur in similar contexts have similar meanings. To use it, represent each word by a list of numbers describing which words appear near it, then compare lists. Words whose lists point the same way, measured by the cosine of the angle between them, are similar. Mikolov and colleagues' word2vec (2013) and Pennington, Socher and Manning's GloVe (2014) made such embeddings dense and learnable, with a few hundred numbers per word instead of one per vocabulary item. Interactive · Who keeps similar company? Thirty invented sentences, with function words dropped and a window of two words either side. Each word's counts are converted to positive pointwise mutual information (how much more often a neighbour appears than chance) and compared with cosine similarity. Everything is computed on the page. Without JavaScript: in the toy corpus the nearest word to “cat” is “mouse” (0.36), then “dog” (0.27); to “bus”, “car” (0.69); to “bread”, “baker” (0.67). Look at the results. “Mouse” and “dog” come out near “cat”, but “baker” comes near “bread”: closeness records association, which mixes similarity of meaning with being the same sort of thing and being found together. Large embeddings capture more, and the famous “king − man + woman ≈ queen” example works for some words. Later analyses (Rogers, Drozd and Li, 2017; Nissim, van Noord and van der Goot, 2020) found that such analogies are less general than the examples suggest, partly because the input words are excluded from the answer, and that embeddings also pick up the biases of their text. Attention: how a transformer looks things up Embeddings give every word one fixed vector, but “bank” in “river bank” and “bank loan” should differ. The 2017 paper Attention Is All You Need (Vaswani and colleagues) built the transformer around a mechanism that rebuilds each word's vector from the words around it. For each position the model forms a query (“what am I looking for?”), and every other position offers a key (“what do I contain?”). The match scores go through a softmax, which turns them into weights adding to 1, and the position's new vector is a blend of the other words' values, weighted by those scores. Stack many such layers and heads, train on text to predict the next word, and you have the architecture behind current large models; for the full picture see How Large Language Models Actually Work (29/02). Here only the lookup matters. weights = softmax( scores ÷ temperature )higher scores get more weight; a high temperature spreads attention out, a low one sharpens it Interactive · What does “it” refer to? … too big… too small Softmax temperature “The trophy would not fit in the suitcase because it was too big / small” is a Winograd-schema pair (Levesque, Davis and Morgenstern, 2012; the idea goes back to Terry Winograd, 1972). The scores are invented to show what a trained attention head might do, not read out of a real model. The softmax is real. Without JavaScript: at temperature 1 with “too big”, “it” attends about 56% to the trophy and 8% to the suitcase; with “too small” the numbers swap. At a temperature of 5 the weights spread out (trophy 14%, others 7–9%). This is why long-range dependencies stopped being hard: the first word can attend directly to the last. What the heads actually compute in trained models is studied (probing work such as Hewitt and Manning, 2019, finds tree-like syntactic structure in internal representations), but a single attention weight is not a full explanation of a model's behaviour. Evaluation: scoring language is hard A system's output needs a grade, and for language there is rarely one right answer. For classification tasks (is this name a person or a place?) the standard measures are precision (of what it flagged, how much was right), recall (of what was there, how much it found), and their harmonic mean, F1. For generation the measures are rougher. BLEU (Papineni, Roukos, Ward and Zhu, 2002) compares a machine translation with human reference translations by counting matching word sequences, with a penalty for too-short output. It made cheap, fast comparison of translation systems possible and it correlates with human judgement over a whole test set. For a single sentence it can be badly wrong. Interactive · A metric looks at words, not meaning Precision and recall of an invented name-finder Confidence needed before it flags a name Simplified BLEU with word and two-word matches (BLEU-2, repeat words counted no more than the reference allows), versus the single reference shown. Real BLEU uses up to four-word matches over many sentences. The name-finder scores 20 invented candidates, 10 of which are real names. Without JavaScript: an identical candidate scores 1.00; a light paraphrase 0.58; reordering the same words 0.91; a paraphrase in entirely different words scores 0 (and so does a string of “the”s); a too-short answer 0.19. At a 0.50 threshold the name-finder has precision 0.64, recall 0.70, F1 0.67. The pattern repeats across the field. Perplexity (how surprised a model is by held-out text) cannot be compared across models with different tokenisers. Human judges disagree, so annotation studies report agreement beyond chance (Cohen's kappa). Benchmarks saturate and leak into training data, so a high score may reflect memorisation. Newer learned metrics (BERTScore, COMET) track meaning better and have their own biases. The skilled habit is to ask what a number would look like if the system were bad in a way the number ignores. What language models do and do not show about language Large language models are neural networks trained on trillions of words to predict the next one. They write fluent text in dozens of languages, produce grammatical sentences with long-distance agreement, and answer questions. What does this show about human language? Careful people disagree, and here are the main positions in their own terms. The models refute, or at least strain, the generative view. Steven Piantadosi (2023) argued that these models learn rich syntax and semantics from input alone, without the innate grammar that Chomsky's tradition proposes, and that a theory of language should be judged partly by how well it supports such models. Usage-based researchers such as Michael Tomasello have long argued that structure emerges from experience, and see the models as consistent with this. The models show little about human language. Chomsky, Ian Roberts and Jeffrey Watumull (New York Times, March 2023) wrote that the models work equally well on possible and impossible languages and so explain nothing about what makes human language what it is. A reply is the study by Kallini and colleagues (2024), which found that models learn artificially "impossible" languages less easily than natural ones,. The sheer amount of data is a second point: the BabyLM challenge limits training to 10 or 100 million words to ask what can be learned at a child's scale, noting that models of the Chinchilla class saw 1.4 trillion words, over ten thousand times what a 13-year-old has heard. If a child learns from far less, the models' success does not settle the poverty-of-the-stimulus argument, whichever side you favour. Form is not meaning. Emily Bender and Alexander Koller (2020) argued that a system trained only on form, with no connection to the world, has no basis for learning meaning in the sense of communicative intent. Their picture of an octopus listening in on a cable became the standard thought experiment. Bender and colleagues (2021) added concerns about scale, environmental and social cost, and calling fluent output understanding (“stochastic parrots”). Defenders such as Christopher Manning have argued that much of meaning is use, so a system trained on use can acquire a good deal of it. Language is not thought. Mahowald, Ivanova, Blank, Kanwisher, Tenenbaum and Fedorenko (2024) distinguish formal competence (rules and patterns of a language), at which the models are strong, from functional competence (reasoning, world knowledge, social understanding), at which they are uneven. They cite evidence from human brains that these two capacities rely on different systems. The point is not that models are failures, but that a fluent sentence is weak evidence of the underlying thought. Where there is agreement: neural networks handle many syntactic patterns once thought to need explicit rules (Linzen, Dupoux and Goldberg, 2016, found this early for LSTM networks), they fail in ways that vary by language, and the data is badly uneven. Of roughly 7,100 languages, a small number get nearly all the attention; Joshi and colleagues (2020) found that most are barely represented in language technology. That is both a scientific gap and a fairness one. Whatever one thinks of the theory, anyone building language technology must choose whom it serves. What a graduate just sees A "word" is a convenient fiction; tokenisers cut text into pieces by frequency, and some languages are cut more kindly than others. Counting already gives a language model, and the long tail of rare words and unseen sequences is what makes it hard. Zipf's law: a few words do most of the work, and most words are rare. Sentences are ambiguous; a parser needs both grammar and knowledge of the world. Embeddings measure association by shared company, which is not the same as meaning. Attention is a weighted lookup: every word rebuilds itself from the words that match it best. Every metric is a proxy; ask what it ignores. Fluent output is not proof of understanding, and failing a test is not proof of its absence; the debate is real. Where it shows up next Psycholinguistics & the Brain (36/28) and Language Acquisition (36/27): models of reading and learning, and the data gap between children and machines. Semantics & Pragmatics (36/24) and Philosophy of Language (36/07): what meaning is, which the models test and sharpen. Machine Learning (34/32): training, overfitting and evaluation, all of which apply directly. In the world: search, spam filters, automatic captions, translation apps, dictation, writing assistants, and the machine-translated page you read in a language you don't speak. If you want to go deeper 📘 Textbooks. Daniel Jurafsky and James H. Martin, Speech and Language Processing (3rd edition, a free online draft that is updated regularly), and Jacob Eisenstein, Natural Language Processing (MIT Press, 2019). Bender's Linguistic Fundamentals for Natural Language Processing (Morgan & Claypool, 2013) is the linguist's companion. 📘 Gentler. Melanie Mitchell, Artificial Intelligence: A Guide for Thinking Humans (2019), a clear and sceptical account of what these systems do. 🌐 Free text. Jurafsky and Martin, Speech and Language Processing, third-edition draft online (the draft dated August 2026 begins with large language models, then covers tokens, n-grams and embeddings). 🌐 Free courses. Stanford CS224N, Natural Language Processing with Deep Learning (slides free; the 2024 lecture videos are on YouTube), and MIT OpenCourseWare 6.864 Advanced Natural Language Processing (Fall 2005, Collins and Barzilay), which covers the classical statistical methods. The takeaway Computational linguistics turned language into things a machine can count, compare and look up: tokens, probabilities, trees, vectors and attention weights. Each step captures something real about language and loses something else. The newest systems write remarkably well, and what that shows about human language, meaning and thought is still an open argument with serious people on every side. Try this: pick a sentence in a language you know and put it through a translation app in both directions; note where the meaning shifts, then change one word in the original and see whether the system notices. Sources & Further Reading Sources. Gage, “A new algorithm for data compression,” C Users Journal 1994; Sennrich, Haddow and Birch, “Neural machine translation of rare words with subword units,” ACL 2016; Petrov et al., “Language model tokenizers introduce unfairness between languages,” NeurIPS 2023, and Ahia et al., “Do all languages cost the same?” EMNLP 2023. Shannon, “A mathematical theory of communication,” 1948. Zipf, Human Behavior and the Principle of Least Effort (1949). Carroll, Alice's Adventures in Wonderland (1865), Project Gutenberg ebook 11 (first 1,235 words used). Chomsky, Syntactic Structures (1957); Pereira, “Formal grammar and information theory: together again?” Philosophical Transactions of the Royal Society A 2000. Marcus, Santorini and Marcinkiewicz, Computational Linguistics 1993; Universal Dependencies, release 2.18 (15 May 2026), universaldependencies.org. Firth, “A synopsis of linguistic theory 1930–1955” (1957); Harris, “Distributional structure” (1954); Mikolov et al. (2013); Pennington, Socher and Manning (2014); Rogers, Drozd and Li, “The (too many) problems of analogical reasoning with word vectors,” 2017; Nissim, van Noord and van der Goot, “Fair is better than sensational,” Computational Linguistics 2020. Vaswani et al., “Attention is all you need,” 2017; Hewitt and Manning, “A structural probe for finding syntax in word representations,” NAACL 2019; Levesque, Davis and Morgenstern, “The Winograd Schema Challenge,” 2012. Papineni et al., “BLEU,” ACL 2002; Cohen's kappa (1960). Piantadosi, “Modern language models refute Chomsky's approach to language” (2023); Chomsky, Roberts and Watumull, “The False Promise of ChatGPT,” New York Times, 8 March 2023; Kallini et al., “Mission: Impossible Language Models,” ACL 2024; BabyLM challenge, babylm.github.io (word budgets and the Chinchilla comparison, read 2 November 2026); Bender and Koller, “Climbing towards NLU,” ACL 2020; Bender, Gebru, McMillan-Major and Shmitchell, “On the dangers of stochastic parrots,” FAccT 2021; Mahowald et al., “Dissociating language and thought in large language models,” Trends in Cognitive Sciences 2024; Linzen, Dupoux and Goldberg, TACL 2016; Joshi et al., “The state and fate of linguistic diversity and inclusion in the NLP world,” ACL 2020. Stanford CS224N and MIT 6.864 pages (read 2 November 2026). The corpus, grammar, scores and candidate translations are invented and computed in the page. 📘 Jurafsky and Martin, Speech and Language Processing 📘 Eisenstein, Natural Language Processing 🌐 Stanford CS224N