Title: Learning semantics and pragmatics from dialogue history

Speaker: Matthew Stone (Rutgers)

Time: 2:15pm-3:30pm, Friday, February 18

Place: Room 4102, CUNY Graduate Center, 365 Fifth Ave (34str&35str).

Abstract:

When two people talk together, they work together to reach a shared
understanding of one another.  This collaborative effort seems to be
crucial to the robustness of natural communication, and to people's
capacity to learn from and adapt to one another's use of language.
 Work on dialogue systems increasingly focuses on improving systems'
skills at interaction and collaboration, with the goal of allowing
machines to talk through new ideas flexibly and successfully.

In this talk, I demonstrate some of the ways such collaborative skills
can improve the functionality of dialogue systems.  In particular, I
describe a system that learns from experience to understand users better
in situated task-oriented dialogue. The system accumulates training
examples for ambiguity resolution by tracking the fates of alternative
interpretations across dialogue, including subsequent clarificatory
episodes initiated collaboratively by the system itself. We realize this
approach by building maximum entropy models over abductive
interpretations in a referential communication task. The resulting model
correctly resolves 81% of ambiguities left unresolved by an initial
handcrafted baseline. A key innovation is that our method draws
exclusively on a system’s own skills and experience and requires no
human annotation.

This is joint work with David DeVault, USC.

Bio:

Matthew Stone is Associate Professor in the Computer Science Department
and the Center for Cognitive Science at Rutgers. He got his PhD in 1998
from the University of Pennsylvania. He studies computational models of
conversation, particularly models of utterance production, for
intelligent agents that interact naturally with human partners. He
serves on the editorial board of the journal Artificial Intelligence and
served as program co-chair for the 2007 North American Association for
Computational Linguistics Human Language Technology Conference (NAACL
HLT).