A Grand Challenge for Computer Science Towards a Testable Theory of Meaning - Finding and Representing the Atomic Elements and the Natural Laws in the Field of Meaning and hence the Internal Structure Of Particular Meanings This paper proposes that computer scientists use the tools available within and adjacent to their discipline to examine and prove that meaning has a structure in the same way that matter has a structure (that is, with constituent elements that combine in lawful ways to give us interesting and useful varieties of new and other meanings). The challenge is to supply a testable theory of meaning; to identify and prove that meaning is not some formless cloud but rather that it has a structure in the same way that matter has a structure and to deploy computer models and simulations of human thought (and / or language) to verify this structure. Although such a theory may have been possible at an earlier time, it is now more likely, and more likely to find supporting evidence, due to recent advances in in vivo neuro-imaging. These recent advances provide both structural information (through functional Magnetic Resonance Imagery, or fMRI) and behavioural, or dynamic information (through event-related brain potentials). The latest research (particularly in these peri-millennial years) in these fields suggest that the brain regions activated during emotional and moral decisions may be specific, identifiable and fairly constant across human populations and also that the evoked brain potentials due to specific words are similar across human subjects to such an extent that the words uttered to the subject may be inferred from the evoked potentials alone (in work that has yet to be replicated in many laboratories but has been repeated several times by the same team). In general terms, as we begin to understand the functions of various identifiable structures in the brain we are simultaneously mapping the role of these structures in different cognitive and semantic contexts. What is required now is a theory that might explain these disparate findings and make predictions based on them. Such theories are possible: one such theory and model is based on arrangements of remembered sensations (and hence the sensory organs) experienced by humans; others rely on the self-organising properties of cell assemblies (in response to targets either internally generated or imposed by the environment); others again focus on a more ‘robotics-driven’ approach that situates intelligence (and hence the creation of meaning) as appropriate (self-motivated) action within a specific environment. Even with these, possibly complementary, viewpoints we have differing, testable theories each with a slightly different focus, namely: sensation, or reflection, or action. This proposal puts forward a belief that the interplay of twenty-first neuroscience and ongoing developments in both neural network algorithms and parallel computing make theories of meaning based on its internal structure both implementable and testable. This challenge is for computer scientists to boldly show that just as, in the physical world, light and time and space are not quite what they appeared to be up until the twentieth century, so meaning may not be the impenetrable mystery that it appears to be. This is a wonderful opportunity for Computer Science.
Proposed Grand Challenge: Towards a Testable Theory of Meaning How does the submission meet the criteria for a Grand Challenge? Does it arise from scientific curiosity about the foundation, the nature or the limits of a scientific discipline? Yes. It does this. In particular it challenges and extends the notion of what science in general could undertake and aspire to, since “meaning” has been assumed to be a matter either exclusively, or else primarily for philosophers and linguists. Will it be obvious how far and when the challenge has been met (or not)? Yes. Realistic tests can be devised that distinguish between different research groups and evaluate in which respects (if not necessarily, how ‘far’ on a unilateral scale) each has made progress. What form might such tests take, in a such a way as to demonstrate which group has made measurable progress and a significant contribution? Possible tests could be along the lines of the following. ‘Based on your ability to represent language-independent, “raw” meaning: • translate this passage into five different natural languages (eg. Japanese, Arabic, Creole, French, Russian) with acceptable performance (or conformance to the interpretation of a competent, different human assessor of each language); represent the meaning of this text, in the same language as the original, without using any of the same words and yet without degradation of meaning; explain the meaning of this poem; explain why this joke is funny.’
• • •
Beyond such questions, a theory should permit and provide some insight into questions that, in the absence of internal structure, might appear to be ill-defined. Or, to put it another way, a “good” theory of meaning should provide answers to difficult questions and hence provide us with insights into the way we think and/or expose misunderstandings between us for what they are – merely misunderstandings. For instance, ‘based on your ability to represent languageindependent, “raw” meaning: • • • • show the internal structure of your concepts of faith, hope and charity; what happens to faith when belief is subtracted? (what, if anything, does it become?) what happens to hope when belief is added? (what, if anything, does it become?) what happens to the idea of harmony when all musical connotation, or sound, is removed? (what, if anything, are you left with?)’.
These example tests are not intended to be biased in favour of any particular theory and should be achieved through a dialogue between groups with differing approaches. A leading theory would be the one that, at any given time, least avoids the tests of others and most provides some answers to such tests.
Is there a clear criterion for the success or failure of the project after fifteen years? As above. Does it give scope for engineering ambition to build something that has never been seen before? Yes. It does this. However, this is by no means a particular motivation for the work that would need to be carried out. A translation or interpretation “device” could conceivably be built out of a theory and model that successfully passes tests such as those proposed above. Does it avoid duplicating evolutionary development of commercial products? For the reason stated above. Does it promise a revolutionary shift in the accepted paradigm of thinking or practice? As suggested at the outset. Does it have the enthusiastic support of the general scientific community? Various scientific communities, including the various neurosciences and cognitive sciences, should be eager to see their work contribute to such an outcome as is proposed. Scientists working within physical science disciplines should be interested to see the methods they have developed for the physical world now being applied to meaning. Does it appeal to the imagination of other scientists and the general public? It must do so. The challenge is readily comprehensible and should fascinate any person interested in how we express ourselves through language. Furthermore, the nature of some of the proposed tests should encourage and embrace the work of poets, philosophers, linguists and cultural theorists and thus encourage a unique dialogue between the arts and the sciences. What kind of benefits to science, industry, or society may be expected from the project, even if it is only partially successful? The dialogue alluded to in the paragraph above would be such a benefit from a partial success. It promises to go beyond what is initially possible, and requires development of understanding, techniques and tools unknown at the start of the project? Although the proposal is motivated by recent, available advances, the further developments required for a solution are not known yet. Does it have international scope? Yes, for sure. For example, many of the approaches to Artifical Intelligence that may be significant in this context, and which stress the action-within-an-environment approach emanate from America (for instance, at MIT) and Japan; indeed, an approach to “Intelligence” that seeks to do away with (compartmentalised) representation and which combines each of the sensation, reflection, and action approaches, as characterized here, has been pursued for some time at the Rockefeller Institute (Synthetic Neural Modelling, Edelman et al). Does it call for collaboration of research teams with diverse skills? Yes it does. The challenge may re-inforce links that already exist or else suggest new collaborations. For instance, strong links between computer science and language research themes are currently active at Middlesex University, and no doubt at many other universities in the UK. Such fruitful collaboration has, on occasion, been formalised in Faculty re-structuring such as happened at Keele University, where a new School of Neuroscience, Computer Science, Mathematics was developed over a decade ago.
How does the project split into sub-tasks or sub-phases, with identifiable goals and criteria, say at five-year intervals? This is not certain. However, given that the tests suggested above may require very different approaches, a modular approach could be taken based on the targeting of individual tests. Can it be promoted by competition between teams with diverse approaches? Yes, by reference to various tests such as those already outlined. In addition, it may well be that teams will ‘compete’ in order to show that chasing the ‘structure of meaning’ is a wrong approach to answering the questions suggested? Is this challenge of some vintage? Not exactly; however, related challenges have been proposed at various times and in various guises and these challenges have remained relevant and motivating of research in computer and cognitive sciences. An example of such a challenge is the Symbol Grounding Problem (Harnad, 1990). This “structure of meaning” approach may thus be seen as a new approach to familiar Artificial Intelligence problems. However, we now have new tools and evidence from computer science and neuroscience to investigate “meaning” and we should now set about providing the computational and theoretical tools to meet the challenge.
Proposer: Yaw Busia, Senior Lecturer, School of Computer Science, Middlesex University. y.busia@mdx.ac.uk