In the previous post we built our linguistic onion - five layers, from phonetics to pragmatics, and we saw how an LLM handles each of them. But after writing that post, one big "but..." lingered in my head.
\nBecause - does an LLM even "understand" what it generates? Does it have some internal model of the world? Does it think?
\nAnd then I stumbled into semiotics. And it turns out that semiotics gives us a brilliant frame for thinking about LLMs - not as an artificial mind, but as a machine of signs. And suddenly everything starts to make sense. Or at least it makes more sense than before ;-)
\nThis is the second post in the "Understanding LLM" series. Today we shift perspective: instead of looking at the layers of language, we look at the very nature of what signs are and how meaning comes into being at all. And why that is key to understanding what an LLM is - and what it is not.
\nBefore we get to LLMs, we need to nail down the basics. Because "semiotics" is one of those words that sounds smart, but what does it actually mean?
\nSemiotics is the study of signs and of how signs create meaning. That's it. Doesn't sound so scary anymore, right? ;-)
\nAnd a "sign" in semiotics is anything that means something. Something that stands for something else. Simple examples:
\nEach of these signs represents something else. And that "representing" is exactly what semiotics studies.
\nTip
\nExperiment: Look around you - how many signs do you see right now? At this moment I see: a WiFi icon (I have internet), a notification on my phone (someone texted), a logo on my coffee mug (a brand). Three signs and I didn't even get up from my chair.
\nIn the previous post we had semantics - the study of the meaning of words and sentences. So how does semiotics differ?
\nIn short:
\n| \n | What it asks | \nExample | \n
|---|---|---|
| Semantics | \n"What does this mean?" | \nWhat does the word "zamek" mean? | \n
| Semiotics | \n"How does this sign even work?" | \nHow is it that a red light MEANS "stop"? | \n
Semantics asks about specific meanings. Semiotics asks about the mechanism of meaning - how it is that anything means anything at all.
\nAnd that is exactly why semiotics is so important for understanding LLMs. Because the question is not "what does an LLM mean", but "how an LLM operates on signs".
\nIn semiotics there are two main traditions you need to know. Two approaches, two ways of thinking about signs. And note - both are important for understanding LLMs, but each from a different angle.
\nSaussure (a Swiss linguist, who lived at the turn of the 19th and 20th centuries) said: a sign consists of two parts.
\n\nAnd the key thing: the relationship between signifier and signified is arbitrary. There is no logical reason why the sequence of letters "c-a-t" means that particular animal. It just... caught on. In Polish it's "kot", in German "Katze", in Japanese "็ซ" (neko) - every language has a different sequence of sounds/letters for the same concept.
\nBut Saussure says something even more important: the meaning of a word arises from its relationship to other words in the system. "Cat" means what it means because it is NOT "dog", it is NOT "house", it is NOT "car". Meaning is differential - it arises from difference.
\n\n\nThis sounds abstract, but in a moment you'll see that this is exactly what embeddings do in an LLM. Really ;-)
\n
Peirce (an American philosopher, a bit earlier than Saussure, but roughly around the same time) had a different approach. For him a sign is not a static pair, but a dynamic process.
\n\nPeirce's triad:
\nAnd that is the key difference: for Saussure a sign is static (a pair), for Peirce it is dynamic, alive, a process. Meaning is not "contained" in the sign - it arises in the process of interpretation.
\nPeirce divided signs into three categories that are super intuitive:
\n| Type | \nDescription | \nExamples | \n
|---|---|---|
| Icon | \nA resemblance between the sign and the object | \nA portrait, a map, the emoji ๐บ (looks a bit like a cat), a folder icon on a computer | \n
| Index | \nA causal or physical connection | \nFootprints in the sand (someone walked here), smoke (fire), a cough (illness), a thermometer (temperature) | \n
| Symbol | \nA convention, a social agreement | \nThe word "cat", a national flag, a red light = stop, the mathematical "=" | \n
Note
\nThe word "symbol" in everyday language means something different than in Peirce's semiotics! In semiotics a symbol is a sign based purely on convention - it doesn't resemble the object (like an icon) and isn't physically tied to it (like an index). The Polish flag doesn't "resemble" Poland and isn't physically connected to it - we simply agreed that these colors stand for that country.
\nTest yourself - what type of sign is this?1
\nDavid Manheim, an AI researcher, used a beautiful metaphor coming from Peirce's semiotics. He called it the "Hall of Mirrors Problem".
\nImagine: you're in a room full of mirrors. You see reflections of reflections of reflections... and there's no window to the outside anywhere. You don't see the real world - you see only... more mirrors.
\nThat is exactly what an LLM does.
\n\nLet's look at this through the lens of Peirce's triad:
\nAn LLM has never seen the world. It has never tasted a strawberry, never touched ice, never heard laughter. It draws all its "knowledge" about the world from text - from what other people wrote about the world.
\nSo when an LLM writes "strawberries are sweet" - it doesn't know they are sweet. It knows that in the texts it was trained on, the word "strawberries" often appears near the word "sweet". That is the difference. And that is precisely the semiotic difference.
\nWarning
\nParadox: An LLM can write a beautiful description of a sunset, even though it has never seen the sun. But it can also write a beautiful description of a sunset on Mars - even though nobody has ever seen a sunset there yet. How does it "know"? From science fiction text. So: signs refer to signs, which refer to signs... a hall of mirrors ;-)
\nOK, now what I promised - let's see how Saussure's theory plays out in code. Because this is genuinely fascinating.
\nRemember what Saussure said? The meaning of a word arises from its relationship to other words. "Cat" means what it means because it isn't "dog", it isn't "house", etc. Meaning is relational.
\nNow think about word embeddings - which we met in the previous post. Each word is represented as a vector in a high-dimensional space. And words with similar meanings are close to each other in that space.
\nThis is exactly Saussure's relational theory of the sign, just implemented in math!
\nfrom gensim .downloader import load \n\nmodel = load ("glove-wiki-gigaword-50" )\n\nking = model ["king" ]\nqueen = model ["queen" ]\nman = model ["man" ]\nwoman = model ["woman" ]\n\nresult = king - man + woman \n\nfrom gensim .models import KeyedVectors \nsimilarities = model .cosine_similarities (result , [queen ])\nprint (f"Similarity to 'queen': { similarities [ 0 ]:.3f } " )It will print something like: Similarity to 'queen': 0.850
But look at this from Saussure's perspective:
\n\nBoth Saussure and Word2Vec say the same thing: meaning is not in the sign itself - it's in its relation to other signs. Saussure came up with this as a theory of language. Engineers at Google came up with Word2Vec as an algorithm. And they reached the same conclusion.
\nWell... almost. Because there's one catch. Saussure assumed that behind the signifier (form) stands the signified (concept). In embeddings we only have a position in space - we have relations, but do we have a "concept"? Is the vector [0.50, 0.68, ...] the concept of "king"?
And that is exactly the question that leads us to the next semiotician...
\nJacques Derrida (a French philosopher, 1960s and 70s) did something bold. He looked at all of Saussure and said: "Wait. And why do you assume speech is more important than writing?"
\nSaussure (and the entire Western philosophical tradition) treated speech as "primary" - closer to thought, closer to meaning. Writing was "derivative" - merely a record of speech, a "sign of a sign". Derrida called this logocentrism - the belief that at the end of the chain of signs there is some "presence", some "thought", some "intention" that bestows meaning.
\nAnd Derrida turned this upside down: writing is not subordinate to speech. Writing is a system in its own right.
\nWhy does this matter for LLMs? Because Elad Vromen, in his paper "Language Models as Semiotic Machines", noticed something brilliant:
\n\n\n\nAn LLM trains on writing (text). It builds a model of writing. It generates new writing. Nowhere in this process is there "speech", "mind", or "intention". Saussure's entire hierarchy - speech > writing - gets inverted. Writing is the only reality an LLM knows.
\n
So when we ask "does an LLM understand language?" - we're asking a bad question. It's like asking "does a book understand what is written in it?" A book doesn't "understand" - but it contains signs that we - the readers - interpret. An LLM is something in between: it isn't a book (because it generates new text), but it isn't a mind either (because it has no access to meanings outside of text).
\nImportant
\nDerrida in a nutshell for LLMs: An LLM doesn't model a "mind" or a "world". An LLM models writing - a system of signs that has its own logic, its own rules, its own coherence. And that writing is enough to generate text that means something to us. But it is us who give it meaning - not the model.
\nThis also explains a phenomenon you've probably noticed: an LLM sometimes says things that are statistically correct, but nonsensical. Because in the system of writing the model operates in, those words fit together well. But we - as humans with access to the world (to objects in Peirce's sense) - see that it makes no sense. The model lacks that "grounding" in reality.
\nDerrida, in his famous essay "Signature Event Context", spoke of the iterability of signs - the fact that a sign can be repeated in a new context and take on new meaning. The word "good" can be a compliment, irony, or part of "good evening" - context changes everything.
\nAnd that is exactly what we see in LLMs: the same prompt in a different context yields a different answer. The model doesn't "understand" context - but statistically it picks up contextual patterns from the training data. So: iterability of signs in its purest, mathematical form.
\nNow we move onto very practical ground. Because if an LLM is a machine of signs, then a prompt - what you type into it - is a semiotic act. Not simply "a command". But an act that creates the frame for meaning.
\nLet's look at the interaction with an LLM through the lens of Peirce's triad:
\n\nSo:
\nNotice: meaning arises only in step 3. The LLM generates a sequence of tokens, but it's you who gives it sense. This is exactly what Umberto Eco meant with his concept of the "open work" - a text doesn't have one fixed meaning. A text is a "frame" that the reader fills with interpretation.
\nTry it yourself. Fire up ChatGPT (or Claude, Gemini - whatever you have) and send these three prompts, each in a new conversation:
\n"Explain what gravity is.""Explain gravity to a five-year-old.""Explain gravity in the style of a Shakespearean sonnet."Each prompt is about the same topic (gravity). But each one changes the semiotic frame - it shifts the tone, the register, the genre, the expected form of the answer.
\n| Prompt | \nWhat changes... | \nSemiotic frame | \n
|---|---|---|
| "Explain gravity" | \n- | \nNeutral, encyclopedic | \n
| "...to a five-year-old" | \nThe audience, the simplicity of language | \nEducational, age-appropriate | \n
| "...in the style of Shakespeare" | \nThe genre, the form, the style | \nLiterary, artistic | \n
This is exactly what Picca calls the "semiotic contract". When you write a prompt, you don't "ask for information" - you set the conditions under which meaning will be constructed. You ask for gravity in Shakespeare mode? You get a hybrid of physics and poetry. That's neither "real" gravity nor "real" Shakespeare - it's a semiotic collage, a new configuration of signs.
\nTip
\nBonus experiment: Try this: "Explain the concept of entropy using metaphors from fairy tales." You'll see how an LLM connects two completely different semiotic zones - physics and fairy tales. That is exactly what semiotics calls translation between cultural codes.
\nOne more concept worth knowing. Juri Lotman, a Russian semiotician, came up with the concept of the semiosphere.
\nThe semiosphere is the space in which signs live. It's an ecology of meanings - a network of cultural codes, genres, discourses, ideologies that interact with one another. Just as the biosphere is the space in which organisms live, the semiosphere is the space in which signs live.
\n\nAn LLM's training data is nothing else but a giant cross-section of the semiosphere. When the model reads Wikipedia, Twitter, books, scientific papers, source code - it absorbs (statistically!) all that diversity of cultural codes.
\nAnd that's why it can write in the style of Shakespeare, translate from German, joke like a comedian, and quote legal regulations - because all of that lives in the semiosphere, and the model has "seen" samples from every zone.
\nBut there's a flip side: the model also absorbs the prejudices, stereotypes, and dominant narratives contained in the semiosphere. Because the semiosphere isn't neutral - it's a cultural space with a history, with power, with ideology. And the LLM, operating on signs from that space, reproduces them in its outputs.
\nWarning
\nWhy does an LLM sometimes talk nonsense? Because the semiosphere is full of contradictions. On one page of the internet "the earth is round", on another "the earth is flat". The model sees both signs and has no access to the "object" (the actual earth) to decide which sign is true. It operates in a hall of mirrors - signs refer to signs, not to reality.
\nTest which semiotician better explains these LLM phenomena. A reminder:
\nWho better explains the fact that...?2
\nHere's our semiotic map in a nutshell:
\n| Semiotician | \nKey concept | \nWhat it tells us about LLMs | \n
|---|---|---|
| Saussure | \nMeaning = relation between signs | \nEmbeddings realize the relational concept of the sign | \n
| Peirce | \nTriad sign-object-interpretant | \nAn LLM has no access to the object, only to signs (hall of mirrors) | \n
| Derrida | \nWriting as a system in itself | \nAn LLM models writing, not mind; training data = the only source | \n
| Lotman | \nSemiosphere - the ecology of signs | \nTraining data is a cross-section of the semiosphere; an LLM navigates its zones | \n
| Eco | \nThe open work | \nAn LLM's output has no single meaning - it requires your interpretation | \n
So: an LLM has no mind. But that doesn't mean it does nothing interesting. An LLM operates on signs - it reconfigures them, connects zones of the semiosphere, creates new configurations of signs. And those new configurations mean something to us - as interpreters.
\nThat is the semiotic answer to "does an LLM understand?". An LLM doesn't "understand" in the human sense. But it generates signs that enter our semiosphere and become part of our process of interpretation. And that process is real, important, and powerful.
\nI know semiotics sounds at first like something from another planet, but I hope you now see how it connects to LLMs.
\nWhich semiotic concept do you find most useful for understanding LLMs? Saussure and his relationality? Peirce and his hall of mirrors? Or maybe Derrida and his "writing"?
\nBecause frankly - I'm still learning this perspective myself. But I feel this is the moment where linguistics, philosophy, and programming meet in one place and create something genuinely fascinating.
\n\n\nWhat's in the next post? We move from theory to practice: How a computer reads text - from counting words to vectors. We'll see how tokenization, TF-IDF, Markov chains, and Word2Vec realize in code what Saussure (meaning is relational!) and Derrida (writing is a system in its own right!) were talking about.
\n
See you next time!
\nSources and interesting links:
\nIf you want to go deeper, here are the materials I used:
\nAnswers: 1) index (causal - temperature causes the liquid to expand), 2) icon (a resemblance to the dog), 3) symbol (pure convention), 4) index (a physical trace of someone who walked here), 5) icon (a resemblance to a person in a wheelchair). โฉ
\nMy answers: 1) Peirce - the model has no access to the object (the experience of poetry), but it generates representamens (text) that we interpret. 2) Saussure - this is pure relationality! A king is defined by what it is not. 3) Peirce - context changes interpretation, and the interpretant creates a new sign. 4) Peirce - the absence of the object in the triad. 5) Peirce + Lotman - the model reproduces signs from the medical zone of the semiosphere without access to the real object (the patient's body). The diagnosis is a sign without grounding. โฉ
\nLLMs (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 โ โ โ\n "not" "happy" makes a noun\nOr in Polish:
\nnieszczฤลliwy = nie + szczฤลliw + y\n โ โ โ\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 | \n็ซใใใใใฎไธใซๅบงใฃใฆใใ | \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") โ 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. โฉ
\n