On February 10, 2021, I will give a talk at the Work In Progress seminar of the CEFISES at UCLouvain, Louvain-la-Neuve. The seminar is open and all interested are invited to join online.
Venue and time
February 10, 2021, 15:00 - 17:00 CET
MS Teams (please email me or the seminar organizers to get a link to our meeting)
What will it take for machines to understand counterfactual conditionals?
Machine understanding is an instance of so called system understanding. We say that a system understands a given domain if it contains appropriate linkages among all (or most) relevant concepts, actions or states. Given our current technology, the systems can easily be programmed to link, e.g., an English word “book” to a French “livre”, but it is much harder to enable a system to correctly link “book” with “writer” or “doctoral student”. A related problem is that of setting up machine understanding in a way which would mimic human understanding. In the context of natural language, a good example of difficulties in implementing human-like understanding emerges from conditionals, especially counterfactual conditionals.
I discuss the problem of formal representation of counterfactual conditionals in a wider context of machine understanding. Two general approaches to machine understanding in artificial intelligence research are connectionism, with machine learning and deep learning as its main paradigms, and symbolism, whose methods are rooted in formal logic. A range of hybrid approaches exists for solving specific problems, such as the implementation of understanding of conditional sentences. In my talk I will discuss how such approaches allow us to derive the notion of understanding, which goes beyond book-livre and ventures into book-writer-type of linkages in the natural language. Thus improved understanding is a step towards approximating human understanding of the natural language.