• Home
  • CLT seminar: Arash Eshghi – Say it and see what happens: learning word meaning from dialogue

CLT seminar: Arash Eshghi – Say it and see what happens: learning word meaning from dialogue


What does it take to bootstrap a language? How is it maintained? How does meaning emerge and become shared? I will review some of the representative literature on the so called Symbol Grounding problem and how it relates to the field of Computational Semantics. I will discuss two prominent approaches to semantics - logical and distributional - and what might be seen as their respective strengths and shortcomings. I will suggest that what most existing approaches lack is sufficient regard for the basic idea that symbols are used to do stuff in the world in interaction with others; and consequently, that their meaning is constrained by (1) their utility in abstract reasoning/planning in a particular task; and (2) semantic coordination pressures in interaction with specific interlocutors.

I then go on to sketch a model of how a structured, probabilistic ontology of object kinds and their properties - represented using the Type Theory with Records framework- can be bootstrapped from dialogue in the context of a simple collaborative referring game. The basic premise in this model is that one can learn the meaning of semantic representations by deploying them in a particular environment and observing their effect - perlocution - to learn what they mean. The model will use a combination of the DS-TTR grammar framework (Dynamic Syntax and Type Theory with Records) for dialogue processing, and Reinforcement Learning for optimisation of what to say. The hope is that one can use such a framework in simulation to test specific hypotheses about what basic mechanisms - e.g. incrementally, repair, rejection, agreement, alignment, etc. - are required for the emergence of shared meaning; but also that it can be used as a basis for developing conversational robots that can learn new concepts and adapt existing ones in interaction with humans.

Short bio:
Arash Eshghi completed his PhD "Uncommon ground: The distribution of dialogue contexts" at Queen Mary University of London in 2009. His research focuses on dialogue and dialogue modelling from a wide range of perspectives including computational methods and experiments on human-human interaction. He is currently working at Heriot Watt University's Interaction Lab on the EPSRC funded BABBLE project: domain-general methods for learning natural spoken dialogue systems, in which speech systems can be trained to interact naturally with humans, much like a child who experiments with new combinations of words to discover their usefulness.

Date: 2015-11-05 10:30 - 12:00

Location: L308, Lennart Torstenssonsgatan 8


add to Outlook/iCal

To the top

Page updated: 2015-10-28 17:24

Send as email
Print page
Show as pdf