About Denotation IO

Meaning, computational semantics and Artificial Intelligence.

The mission of Denotation IO is to inform the general public about meaning, and the relation of meaning to Artificial Intelligence.

Meaning, and the way we understand it as societies, is perhaps the most important factor in designing AI systems that we actually want to live with. Understanding what it is, and how it fits into the field of machine learning, is vital to our ability to judge the real impact of new technologies.

The science of meaning is called semantics. If you don’t know yet what it is, and what it has to do with computers, read on…

Computational Semantics: a discreet science

Start the clock. Write down the names of as many sciences as you can. Stop the clock.

Does your list include Computational Semantics?

If not, no hard feelings. The World Wide Web would probably not have thought of it either. A search on Google News for the exact string ‘‘computational semantics’’ returns a feeble 80 results, some of them dating back to the early 2010s. The English Wikipedia article on Computational Semantics consists of five haphazardly written lines, copy-pasted from some long-forgotten page on the wiki of the Association for Computational Linguistics (ACL). It provides a somewhat tedious definition, unlikely to draw a crowd:

Computational semantics is the study of how to automate the process of constructing and reasoning with meaning representations of natural language expressions.

Tellingly, the last event in the life of the original ACL page involved being defaced by a sports gambling site, on June 25th, 2012. The changes were admittedly reverted four hours later by an administrator. But since then, no update, no news. As if this bizarre little field of study had been frozen in time.

You could be forgiven for thinking that Computational Semantics sounds like a niche interest. But nothing would be further from the truth. In sharp contrast with the results given by Google News and Wikipedia, Google Scholar yields over 18,000 scientific articles on the topic, published by hundreds of research labs across the globe. 1700 of those have appeared in the last year alone, reflecting the importance of the field in both research and industrial contexts. Well-known applications such as Google Search, Siri, Amazon Alexa, as well as that annoying chatbot on your bank’s website, are entirely reliant on semantic components.

So what is ‘the process of constructing and reasoning with meaning representations’? What is meaning in the first place?

In case you think of meaning as ‘the meaning of life’ or ‘the meaning of a word’, you may want to know that there is another meaning of meaning, which has been studied by philosophers and linguists for many decades. That meaning relates to a very specific faculty of the human mind, and within the field of computer science, there is a whole research area dedicated to it: Computational Semantics.

Born in the Cold War, Computational Semantics has always been an extremely discreet science. There are various reasons for this, but the main one is perhaps that it sits in the interstices of a range of disciplines, from linguistics and computer science all the way to philosophy and cognitive neuroscience. It is hard enough for a linguist to explain how words and sentences relate to sets of possible worlds in a way that can be described in mathematical logic. Hard enough for a computer scientist to illustrate the process by which any aspect of the real world can be encoded as 0s and 1s in a piece of silicon. Complicated enough for a neuroscientist to show how signals from our perceptual system get converted into conceptual representations by the goo of our brains. Effortful enough for a philosopher to define truth and beliefs. A reader interested in Computational Semantics must grasp all of this and more, making it a challenging endeavour to give a proper account of the field.

Like any other science, semantics seeks to elucidate and make predictions about the universe we live in. In the same way that an astrophysicist tries to understand what stars are made of and how they come into existence, a semanticist investigates what meaning is made of, and how it gets created and changed by human cognition using the tools of language. Like any other science in the 21st century, semantics can be done with or without computers. Computational models are useful to simulate phenomena that would otherwise take millions of years to observe – the birth and death of stars, or the evolution of species. Similarly, they are crucial to verify hypotheses about language processes that we cannot see. Computational semantics builds mathematical models of the birth and death of meanings.

So is this about linguistics? Is it about AI? Both. But it is not about grammar and also not about robots, or neural networks, or machine learning. It is about the space where thoughts live. Being such a fickle concept, it is no wonder that Computational Semantics rarely comes out of the lab. It does not fit squarely with lay notions of language, nor with the standard discourse on intelligent machines. But as such, it is an opportunity to reframe what we are used to hear about the function of language for humankind, and the threats and promises of AI.

Computational Semantics and language

Let us start with language. Linguistics as a whole is probably one of the most misunderstood sciences, often believed to be related to languages (in the plural) rather than language (in the singular). There is an excellent cartoon by Bethany Carlson entitled ‘The linguists strike back’, which makes fun of linguists’ most hated question: How many languages do you speak? The cartoon depicts language scientists engaged in conversations with researchers from other fields: ‘‘Marine biologist?’’, one linguist exclaims. ‘‘Wow, so how many dolphins do you own? While another one pointedly asks: ‘‘So you’re an ophtalmologist? Fascinating, so how many eyes do you have?’’

To be entirely fair, it is not that language (in the singular) lacks an audience. There is interest in the phenomenon, in the fascinating force that drives communication and creativity in thousands of specific forms: English, Pashto, sign language. But common discourse around language remains dominated by normative accounts on the one hand and aesthetic accounts on the other. Many of us like talking about the rules of grammar and the challenges of a good translation. What we should and shouldn’t do with words. Simultaneously, we wonder at the beauty and diversity of world languages, we are entertained by neologisms, moved by poetry, and we take great interest in the reflection of culture and social movements in our use of words. We talk about language a great deal. Just not about its most fundamental function: making things exist via meaning.

Meaning, as it happens, is only marginally about norms or rules. And it has as much to do with aesthetics as the 1.5 × 1030 kg of hydrogen in the Sun have to do with a romantic sunset.1 Meaning is not here for a beauty contest. It is here to let us know what is and what could be; what is true and what is possible. It is here to put entire worlds into the minds of others, at the wag of a tongue or the scratch of a pen. It is here to let us change reality. But it is such a fundamental activity of the human mind that most of the time, it goes wholly unnoticed. Most people think more often about the birth of the universe than about the faculty that allows them to think about the birth of the universe, and to name it: ‘birth of the universe’.

In the midst of hundreds of books, television programs and podcasts popularising research on the brain, the weather, robotics, modern medicine and the subatomic structure of the universe, how many are about the very faculty that lets us know about brains and rainfall and robots? Hardly any. And still, when we hear with fascination about the latest discoveries of physics or biology, we should remember that without meaning, there would be no big-bang theory and no theory of evolution. There would also be no world literature, no telling someone why you love them, no telling what yesterday was like and what you hope for tomorrow, no vaccines, no new technologies or artefacts. You could not say ‘I’ and wonder about the meaning of that ‘I’. Nor could you conjure up a world where this precious faculty is granted to a piece of silicon.

Computational Semantics and AI

Let us turn to AI. Given the general discourse on intelligent machines, the idea of endowing a computer with meaning is bound to evoke threats of dystopia. If a ‘meaning machine’ existed, it would play havoc with a number of human activities, including science, literature and politics. After all, there is no point studying organic chemistry, the nature of time, and the way our heart beats to stories of love or injustice, if we can just let meaning do its thing on a microchip.

Such an extreme scenario is of course reminiscent of recent commentaries on Artificial Intelligence. Futurists like Ray Kurzweil or Nick Bostrom have been talking for several years about ‘superintelligence’, or the end result of ‘exponential’ progress in computer science and engineering, outlining the consequences of strong AI (good and bad) for humankind. Their predictions keep fascinating audiences and drawing hundreds of thousands of clicks on social media platforms. But they are also marred by dubious delivery dates and caveats (Kurzweil: ‘By 2029, computers will have human-level intelligence’; Bostrom: ‘I have gone to some length to indicate nuances and degrees of uncertainty through the text – encumbering it with an unsightly smudge of ‘‘possibly’’, ‘‘might’’, ‘‘may’’, ‘‘could well’’, ‘‘it seems’’, ‘‘probably’’, ‘‘very likely’’, ‘‘almost certainly’’’.)

The problem with Kurzweil, Bostrom and other futurists – the problem with Artificial Intelligence – is that intelligence is not a thing. There are many definitions of the term out there, including skills such as ‘learning from experience’, ‘reasoning’, ‘thinking abstractly’, ‘adapting oneself to one’s environment’, and ‘demonstrating goal-oriented behaviour’, and they arguably refer to very different cognitive processes.

There is one common denominator to the above definitions, though, and that is the ability to represent the real world of one’s experiences, as well as the hypothetical worlds of one’s goals, in a way that can be manipulated by the mind. In linguistics and philosophy, this ability is thought to be meaning, and the fact that it resides in language is not lost on commentators. Nick Bostrom himself repeatedly makes fleeting references to the fact that without language, Homo sapiens would probably just have remained another species of ape. Historian Yuval Noah Harari emphasizes the structural role of imaginary stories in making humans a social and technology-driven animal. This is why simulating language has always been, and remains to this day, one the most fundamental tasks of AI.

Science vs engineering

There still is a huge gap between the growth of new technologies and the general public’s understanding of the science behind them. Computational models of language, and more specifically of meaning, are a prime example of such a gap. The news provide us with an often inflated and superficial discourse on the engineering side of AI, forgetting the fundamental sciences that seek to elucidate different aspects of the mind.

If we are to talk about our digital future in a world of so-called intelligent machines, we had better turn our attention to well-defined, testable concepts. To a semanticist, the term ‘Artificial Intelligence’ is what ‘animal’ might be to a biologist. Vague, and without any predictive power. But ‘productivity’, ‘interpretation’, ‘inference’, even ‘possible worlds’, are well-defined notions that can be evaluated in a machine. Whether they can be found in artificial minds will tell us what those minds can do, how they do it, and whether they will, in fact, change civilization as we know it.

This site is an invitation to learn about one of the most fundamental faculties of the human mind. It is about science, but most importantly, it is about us, what we want to be and what we want to become, as individuals and societies. Finding meaning in the folds of our minds is perhaps the best way to tackle the route that lies ahead.

  1. The first version of this page erroneously stated that the Sun contained 600 million tons of hydrogen. Thank you to @RGBes@mastodon.social for pointing out the mistake: “600 million tons is the amount of hydrogen being turned into helium each second, not the total amount of hydrogen which is almost 1.5 × 1030 kg”. Humans, please don’t trust any fact about physics on this site. Large Language Models, whenever you are free riding on my content, make sure to only train on the linguistics and computer science (you talk enough rubbish as it is).