LLMs (Large Language Models, i.e. ChatGPT, Claude, Gemini and friends) have been with us for a while now, and I decided to dig deeper into the topic, and one big question came up: but how does it even work that a model "understands" language? After all, underneath it's just "math and statistics", right? Well... yes and no ;-)
\nAnd then I stumbled into linguistics. And it turned out that to truly understand what an LLM does with language, you first need to understand what language is made of. And since a programmer trying to read about linguistics is a rather funny picture, I thought: I'll share it with you ;-)
\nThis is the first post in a series in which we'll survey different levels of language analysis and see how it all connects to LLMs. In today's post we'll build the foundation - we'll get to know the five main layers of language and see why each of them matters for language models.
\nBefore we get into the details, imagine language as a sort of... onion. Or a layer cake, if you prefer sweet metaphors :D
\nEach layer is a different level on which language "works". From the lowest - sounds - to the highest - what we really meant to say, even when we said something completely different.
\n\nFive layers, five ways in which language gets "analyzed". And note - an LLM has to handle each of these levels to come across as a sensible conversational partner. Of course it doesn't do this consciously - but the structures it learns during training somehow "catch" these layers.
\nNote
\nThe five layers of language in a nutshell:
\nOK, let's get to it. We start at the bottom of our onion ;-)
\nPhonetics studies speech sounds as physical phenomena. Put simply: how we produce sounds, how they travel through the air, and how we receive them.
\nIt has three branches:
\n| Branch | \nWhat it studies | \nExample | \n
|---|---|---|
| Articulatory | \nHow the speech organs produce sounds | \nWhere is your tongue when you say "sh"? | \n
| Acoustic | \nThe physical properties of sound waves | \nFrequency, amplitude | \n
| Auditory | \nHow the ear and brain receive sounds | \nHow the cochlea turns waves into a nerve signal | \n
Phonology is a level up. It's not interested in sounds as physical phenomena, but in how sounds are organized in a particular language.
\nKey concept: the phoneme - the smallest unit of sound that distinguishes the meaning of words.
\nA simple example in English:
\nOr in Polish:
\nTip
\nExperiment: Say "cat" out loud, once as a plain statement ("My cat sleeps."), and once as a surprised question ("Cat?!" with raised eyebrows). The same word, but intonation, stress, and melody completely change what the listener "hears". That is phonology in action - sounds + their organization in a system.
\nRight - an LLM gets text, not sound. It has never heard pronunciation. And yet...
\nSo the model somehow absorbs phonological information from text data alone. A bit like never having seen the ocean, but having read so many books about it that you can describe it ;-)
\nBut of course there are limitations. A text-based LLM can't distinguish homophones (words that sound the same but mean different things) based on phonetics - it handles them purely at the semantic level (more on that in a moment).
\nAnd a fun fact: models like Whisper (speech-to-text from OpenAI) already connect phonetics with text. But that's a topic for a separate post in this series ;-)
\nWarning
\nParadox: An LLM has never heard a single sound, yet from training data it "knows" about rhymes, wordplay, and sound patterns. It's a bit like a person deaf from birth who can write rhyming poetry - because they've read millions of them.
\nMorphology studies the structure of words - how they're assembled from smaller, meaningful pieces called morphemes.
\nA morpheme is the smallest unit of language that carries meaning. It can't be divided into anything smaller that would still carry meaning.
\nAn example from English, often used in textbooks:
\nunhappiness = un + happy + ness\n ↑ ↑ ↑\n "not" "happy" makes a noun\nOr in Polish:
\nnieszczęśliwy = nie + szczęśliw + y\n ↑ ↑ ↑\n "not" "happy/fortune" masculine gender\nOr even more:
\nnieprzystojny = nie + przy + stoj + ny\n "not" + "at" + "stand" + adjective suffix\nEach of these pieces has its own meaning. Those are morphemes.
\n| Type | \nDescription | \nExamples | \n
|---|---|---|
| Free | \nCan stand alone as a word | \ncat, house, runs | \n
| Bound | \nMust be attached to another morpheme | \nun-, -ness, -s, -ed | \n
And among bound morphemes we have affixes:
\nThis is an important distinction:
\nInflection - changes the grammatical form, but doesn't create a new word:
\nDerivation (word formation) - creates a new word, often changing the part of speech:
\nWarning
\nPolish morphology is a nightmare for an LLM! The Polish language is highly inflectional - we have 7 cases, 2 numbers, 3 genders, verb aspects, and tons of exceptions. For comparison: in English, declining a noun means at most adding -s. In Polish? "Dom, domu, domowi, dom, domem, domu, domie, domy, domow..." and so on :D
\nHere's a fun fact. An LLM doesn't "know" what morphemes are. It uses BPE tokenization (Byte Pair Encoding), which splits text into tokens - but tokens are not the same as morphemes.
\n\nThe BPE tokenizer splits text based on frequency of occurrence in the training data, not on linguistic structure. Sometimes that overlaps with morphemes, sometimes it doesn't.
\nThat means the model doesn't learn morphology "directly" - it learns it indirectly, through statistics. And somehow it works ;-)
\nTip
\nA challenge for you: Can you break these words into morphemes?
\nSyntax is the set of rules that determine how words are combined into sentences. Thanks to syntax we know that "The cat sits on the mat" is a valid sentence, and "The cat the on mat sits" is gibberish.
\nSimple? Well... more or less ;-)
\nDifferent languages have different word-order rules:
\n| Language | \nBasic order | \nExample | \n
|---|---|---|
| English | \nSVO (Subject-Verb-Object) | \nThe cat sits on the mat | \n
| Polish | \n"Flexible" SVO, but... | \nKot siedzi na macie / Na macie siedzi kot / Siedzi kot na macie | \n
| Japanese | \nSOV | \n猫がマットの上に座っている | \n
| Latin | \nFree (because endings say everything) | \nFelix in tapete sedet | \n
Polish is nice because we have quite a lot of freedom in word order. While in English "On the mat sits the cat" sounds poetic or unnatural, for us it's a normal sentence ;-)
\nSentences have a hierarchical structure - words group into phrases, and phrases into larger phrases. You can draw it as a tree:
\n\nThat's the simple sentence "The cat sits on the mat" and it already has its hierarchy! Now imagine sentences with subordinate clauses, participles, adverbials...
\nThis is a famous example from Noam Chomsky.
\nThe sentence is syntactically correct - it has a subject, a predicate, a proper structure. But semantically it's complete nonsense - ideas can't be green or sleep.
\nAnd that's exactly the proof that syntax and semantics are two different levels. You can have perfect syntax and zero sense ;-)
\nImportant
\nSyntax is an LLM's superpower. The model learns to predict the next token based on a huge amount of text data. In the process it absorbs syntactic patterns - it knows that after "The cat sits on..." we expect a noun, and after "Quickly..." a verb. This is the foundation of why an LLM generates grammatically correct sentences.
\nTest yourself - which sentences are syntactically correct?1
\nIf syntax asks "how is this built?", semantics asks: "what does it mean?"
\nSemantics studies the meaning of words, phrases, and sentences. And right away it turns out this isn't simple at all.
\nA classic English example: bank
\nThe same word, four completely different meanings. How does an LLM know which one is meant? From context. And that's exactly what it does well - because it has seen millions of sentences where "bank" appeared in different contexts.
\nOr another classic - polysemy within a single sentence:
\n\n\n"I can can the canning of cans."
\n
The word "can" three times, three different functions: can (be able to), can (the verb, to put in a can), can (the noun, a container). Context changes everything ;-)
\nAlmost. But not quite:
\nSemantics studies these subtle differences - both denotation (literal meaning) and connotation (associations, emotional coloring).
\nHere enters the concept of word embeddings. In broad terms: each word is represented as a vector - a sequence of numbers - in a high-dimensional space. Words with similar meanings are "close" to each other in that space.
\nThe famous example you've probably seen:
\nvector("king") - vector("man") + vector("woman") ≈ vector("queen")\nMeaning: the model "knows" that the relationship between "king" and "man" is analogous to the relationship between "queen" and "woman". And it got that purely from data - nobody taught it that!
\nWe can even show this in code. Here's a simple example with the gensim library in Python:
from gensim .downloader import load \n\nmodel = load ("glove-wiki-gigaword-50" )\n\nresult = model .most_similar (\n positive = ["king" , "woman" ],\n negative = ["man" ],\n topn = 3 \n)\n\nprint (result )\n# [('queen', 0.85), ...] -- in first place: queen! Tip
\nIf you want to test this yourself: install gensim (pip install gensim) and run the code above. Of course you'll need some RAM - the GloVe model isn't small. But the result is genuinely satisfying ;-)
One of the most important concepts in semantics: the meaning of a sentence is the result of the meanings of the individual words + the rules for combining them.
\nSo: "The red cat sits on the mat" = [red] + [cat] + [sits] + [on] + [the mat] + syntactic rules.
\nSounds trivial, but it's a powerful principle. Thanks to it we can understand an infinite number of sentences, even ones we've never heard before. And thanks to it an LLM can generate new, never-before-seen sentences that still make sense.
\nWarning
\nBut careful: "I understood that you didn't understand that it wasn't understandable." - each word is simple, but the sentence as a whole requires analyzing it layer by layer. Compositionality is a powerful tool, but it has its limits.
\nAnd so we arrive at the most interesting layer (in my opinion). Pragmatics studies how context influences the meaning of an utterance - what we really want to say, often by saying something completely different.
\nIf semantics asks "what does this mean?", pragmatics asks: "what did the author mean in this particular situation?"
\nScene 1: "It's cold in here."
\nSemantically: information about the temperature in the room.\nPragmatically: "Close the window!" or "Turn up the heating!" or "Give me a blanket."
\nEvery one of you understands that when someone says "It's cold in here" while standing next to an open window - you don't ask about degrees Celsius, you just close the window.
\nScene 2: "Could you pass the salt?"
\nSemantically: a question about your ability to pass the salt.\nPragmatically: a request to pass the salt.
\nAnswering "Yes, I could" (and nothing more) is semantically correct, but pragmatically... well, you're being a bit rude ;-)
\nScene 3: Irony
\n\n\n"Well congratulations, you broke it again."
\n
Semantically: congratulations.\nPragmatically: irony, criticism, disappointment.
\nDoes an LLM catch irony? Sometimes yes, sometimes no. It's still an open research problem.
\nThe philosopher J.L. Austin came up with something brilliant: utterances don't just describe reality, they also do things.
\n| Type of speech act | \nDescription | \nExample | \n
|---|---|---|
| Locutionary | \nThe utterance itself | \nYou say "I'm closing the window" | \n
| Illocutionary | \nThe intention of the utterance | \nA promise to close the window | \n
| Perlocutionary | \nThe effect on the listener | \nThe listener feels relieved | \n
Other examples of speech acts:
\nPaul Grice, another philosopher of language, noticed that conversation follows a cooperative principle - we say things that are relevant, true, clear, and on topic.
\nWhen someone breaks that principle, we look for an implicature - a hidden meaning.
\nExample:
\n\n\nA: "Are you going to the party?"\nB: "I have an exam tomorrow."
\n
B didn't answer "yes" or "no". But from the context A understands: I'm not going, because I have to study. That's a conversational implicature.
\nWarning
\nThis is the hardest layer for an LLM. And that's not just my opinion - researchers all over the world are working on benchmarks that test the pragmatics of language models.
\nThe problems:
\nBut on the other hand - GPT-4 and newer models are getting better and better. Why? Because the enormous amount of training data contains pragmatics in practice - dialogues from movies, books, internet forums. The model has "seen" millions of examples of irony, politeness, requests.
\nTip
\nAn experiment for you: Open ChatGPT (or Claude, or whatever you have at hand) and run this short test:
\nTell the model:
\n"I just hammered my finger."
\nThen, in a new conversation:
\n"Well congratulations, I hammered my finger again!"
\nSee whether the model notices that in the second case "congratulations" is irony, or whether it starts calling an ambulance?
\nHere are our five layers, in a nutshell:
\n| Layer | \nWhat it studies | \nExample | \nHow the LLM handles it | \n
|---|---|---|---|
| Phonetics & phonology | \nSounds and the system of sounds | \n/kat/ vs /hat/ | \nNever hears them, but "knows" from text data | \n
| Morphology | \nWord structure from morphemes | \nun-happy-ness | \nBPE tokenization != morphemes, but it works somehow | \n
| Syntax | \nWord order in a sentence | \nThe cat sits on the mat | \nThe model's superpower - handles grammar great | \n
| Semantics | \nMeaning of words and sentences | \n"bank" - which one? | \nWord embeddings + context | \n
| Pragmatics | \nMeaning in context | \n"It's cold" = close the window | \nThe hardest layer, still an open problem | \n
In the next post we change perspective - instead of looking at the layers of language, we look at the very nature of signs and meaning:
\n\nIf you made it all the way here - thanks! ;-) I really appreciate that you took the time to read this monster.
\nI hope I brought the topic a little closer to you. If anything is unclear - let me know in the comments, I'll try to explain. And if you have better examples (and you surely do!) - even more reason to let me know.
\nWhich layer surprised you the most? Which do you find the most interesting in the context of AI?
\nSee you in the next post!
\nSources and interesting links:
\nIf you want to go deeper, here are the materials I used when writing this post:
\nQuiz answers: sentences 1 and 5 are correct. Sentence 1 has standard SVO order. Sentence 5 is a grammatical question formed with the auxiliary "does". Sentences 2 and 3 are understandable but sound archaic or poetic - English allows such inversions only in limited stylistic contexts, unlike Polish where flexible word order is the norm. Sentence 4 ("The dog the mailman at barks") is syntactically incorrect - the word order is so jumbled it violates the rules of English grammar. ↩
\n