Getting meanings out of, and back into, natural language, computationally.

6^th December 2016, 13:00 Ashton Lecture Theater
Dr. Andrew Gargett
Hartree Centre,
Science and Technology Facilities Council

Abstract

The meanings we exchange in our everyday interactions with each other, largely arise through the language we use. Given the complexity of such phenomena, Natural Language Processing has typically focused on more readily available evidence for such meanings, leading to a series of very interesting and fruitful techniques for modelling meaning through text alone; consider the ongoing success of vector-space models of word meanings, where essentially the meaning of a word is determined by other words it typically occurs with. But other aspects of our everyday linguistic interactions points to there being far more than this to natural language meaning; we talk about feeling "down" or "up", or indeed "blue" or "rosy", depending on our moods, how we can "catch" colds, or "throw" a party, that yesterday is "behind" and tomorrow "in front", we are "in" or "out" of a particular group, or indeed can be "in" or "out" of love, someone else's favour, or even our own minds. Much of our linguistic interaction involves such phenomena, pointing to the need for richer representations of the world and not only of words, to be able to model what is being exchanged in and through such meanings. In this talk, I will provide examples of work I have done and am currently doing, in both natural language understanding and generation, on how language is "grounded" in our surroundings, through modelling such phenomena as figurative forms of language (in particular, metaphor), as well as language used to express perceptual and affective meanings. I will discuss how accounts of perceptual experience and emotions across the cognitive sciences can enrich our models of linguistic meaning, and how such work can help extend and improve computational approaches to natural language semantics.