For centuries, theories of meaning have been almost exclusively interested in philosopher, discussed in seminar rooms and at conferences for small special audiences.
But the emergence of large language models (LLMs) and other “Foundation Models” have changed it. Suddenly, mainstream media live with speculation about what is only trained to predict the next word in a order can really understand the world.
Skepticism occurs of course. How can a machine that generates language in such a mechanical way, grease ‘meanings? Just treatment of text, but fluently, does not seem to suggest any kind of deeper understanding.
This kind of skepticism has long history. In 1980, the philosopher John Searle suggested a thought experiment known as the Chinese space where a person who does not know Chinesse follows a set of rules to manipulate Chinese characters and produce Chinese resorts for Chinese questions. The experiment is intended to show that since the person in space never understands the language, symbolic manipulation alone cannot lead to semantic understanding.
Similarly, today’s critics often claim that since LLMs are only able to treat “form” – symbols or words – they can in principle gain understanding. Meaning depends on the relationship between form (linguistic expression or sequences of tokens in a language model) and something external, theses that critics claim, and models that are only trained in form learn nothing about these conditions.
But is that true? In this essay, we will argue that language models cannot only go back meanings.
Probability room
At Amazon Web Services (AWS) we have examined concrete ways of characterizing meaning as represented by LLMS. The first challenge with these models is that there is no candidate for “where” meanings could stay. Today’s LLMs are only used models for decoder; Unlike models that are only codes or codes decoder, they do not use a vector space to resume data. Instead, they are used in a distributed way across the many layers and attention heads for a transformation model. How should we think about meant representation in such models?
In our paper “Meaning Representations from Bane of Authorous Models” we offer an answer to this question. For a given sentence, we consider the probability distribution of all sorts of sequences of tokens that can follow it, and the set of all such distributions defines a representative space.
To the extent that two sentences have similar continuation problems – or Banes – The closer together in the representative space; To the extent that their probability distributions are different, they are further apart. Says that produce the same distribution of continuations are “equivalent” and together they define a Class equivalence. The meaning of a phrase is then the equivalent class to which it belongs.
Within Natural Language Processing (NLP), it is widely recognized that distribution of words in language is closely linked to their meaning. This idea is Nown as the “distribution hypothesis” and is often invoked in the context of methods such as Word Word2VEC Marriages, which build opinion representations from statistics on the word co-occé Jourpe. But we think we are the first to use the distribution even as the primary way of making sense. This is possible as LLMs offered way to evaluate these distributions of calculation.
Of race, the possible continuations of a single sentence are effectively endless, so even with the help of an LLM we can never quite describe their distribution. But this impossible reflects the basics Undrecognation of meaningThat applies to both humans and AI models. Meanings are not observed directly: They are coded in billions of synaps in a brain or billions of activations of a trained model that may be to make expressions. Any final number of expressions can be compatible with several (actually infinite mary) meanings; Which one means the human – or the language model – interest To communicate can never be nown with certainty.
What is surprising, however, is that, despite the great dimensality of today’s models, we do not have to try billions or trillion of the trajectory to characterize a meaning. In a handful – say, 10 or 20 – is sufficient. Again, this consists of human linguistic practice. A teacher asked what a particular station means that it will typically rephrase it in a few ways, on what could be described as an exercise of identifying the equivalent class to which the stern belongs.
In experiments reported in our paper, we showed that a measure of beds that uses LLMS off the shelf to try token lane largely matches human comments. In fact, our strategy surpasses all competing approaches on zero-shot benchmarks for semantic textual equality (STS).
Form and content
Does this suggest that our paper’s definition of meaning – a distribution over possible course – reflects what people do when the Ascibe meaning? Again, skeptics would say that it might not be able to: Text certificates are based only on “form” and lacks the external grouping that needs meaning.
But probability of continuations can capture something deeper about how we interpret the world. Consider a sentence that begins “on top of the dresser stood …” and the probability of this possible sequel to this phrase: (1) “a photo”; (2) “an Oscar -Statuette”; And (3) “an entrance of plutonium”. Don’t these probabilities tell you anything about what you can actually expect to find on Somon’s dressing? The probabilities of all sorts of sentences can be a good guide to the likelihood of finding different objects on top of teachers; In this case, the “formal” patterns coded by LLM would tell you something special about the world.
However, the skeptic may answer that it is mapping words to objects that make the words sense, and the mapping is not inherent to the words themselves; It requires human interpretation or other other mechanism that is external to LLM.
But how do people do that mapping? What happens inside you when you read the phrase “The Obskks on the top of the dress”? Maybe you have obtained something that somehow feels indefinite – an overlay of the dress, seen from multiple angles or heights, says, with abstract objects in a certain range of sizes and colors on top. Maybe you also include the possible rent of dress in the room, the other wandering of the room, the feeling of the dress’s tree, the smell of dress or objects on top of it and so on.
All of these options can be captured by probability distribution, over data in several sensory modalities and in several conceptual schedules. So maybe meaning for people also means probability of continuations, but in a multisensory space -resistant of a textual space. And in that view, when an LLM calculation of token sequences, it gets access to meaning in a way similar to what people do just in a more limited space.
Skeptics may claim that the passage from the multisensory world to implement language is a bottleneck that the meaning cannot push through. But this passing could also be interpreted as a simple projection similar to the projection from a three-dimensional scene down to a two-dimensional image. The two-dimensional image only provides partial information, but in many situations the scene remains quite understandable. And since language is our hand tools to communicate our multisensory experiences, the projection for text may not be so “lost” after all.
This does not mean that today’s LLMS seized meanings in the same way that humans do. Our work only shows that large language models development of internal representations with semantic value. We have also found evidence that such representations are composed of discreet entities that concern each other in complex ways – not just closeness but directional determination, courses and inclusion.
But these structural conditions can be different from the structural conditions in the languages used to train the models. It would remain true, even if we trained the model on sensory signals: We do not directly see what significance suppresses a certain expression, for a model more than for a human being.
If the model and man have been exposed to similar data, how, and if they have shared enough experiences (today, the annotation is the media to share), there is a basis on which to communicate. Adjustment can then be seen as the process of translating between the model is the “inner language” – we call it “neurals” – and natural language.
How faithful can that customization be? As we continue to improve these models, we will have to meet the fact that even people lack a stable, universal system of common meanings. LLMS, with their separate approach to the processing of information, can simply be another voice in a diverse chorus of interpretations.
In some form, questions about the relationship between the world and its representation have been central to philosophy for at least 400 years, and no final answer has appeared. As we move towards a future where LLMs are likely to play a bright and bigger role, we should not reject ideas that are based only when intuitions go continuously to ask these difficult questions. The apparent restrictions of LLMs may be just a reflection of our poor understanding of what the meaning actually is.