NLP@Pitt

"Reinforcement Learning Of Dialogue Strategies Using Restricted Contexts"

by Matthew Frampton (visiting PhD student from University of Edinburgh) This thesis investigates restricted contexts in reinforcement learning (RL) of effective dialogue strategies for information seeking spoken dialogue systems (e.g. COMMUNICATOR Walker et al. 2001). The contexts used are richer than in previous research e.g. Levin and Pieraccini 1997,Scheffler and Young 2001,Singh et al. 2002,Pietquin and Renals 2002 which use only slot-based information i.e. whether or not a slot (e.g. destination city) has been filled and the confidence score associated with any supplied value. The contexts remain much less complex than the full dialogue ``Information States'' explored in Henderson et al. 2005, for which tractability is an issue. Feature engineering is used in order to identify relevant context features from the COMMUNICATOR data. The reinforcement learner then uses a context which includes these relevant features plus the slot-status features. It learns dialogue strategies as it interacts with n-gram user simulations, the probabilities for which are derived from the COMMUNICATOR data. Finally, the learned strategies are evaluated using human subjects. The central hypothesis is that the additional contextual information will enable the learning of more effective dialogue strategies. Results have been obtained in experiments which use n-gram user simulations in both training and testing. The baseline strategy was learned with only the slot-status information. The best performing strategy was learned after adding both the last system and user dialogue moves. It improved over the baseline by 7.8% in average reward per dialogue (significance level p < 0.005) and over the (hand-coded) COMMUNICATOR systems, by 65.9%. The new `emergent' strategies do better in problematic situations where they employ `focus switching' and make effective use of the `give-help' action.

Manual Annotation of Opinion Categories in Meeting

Speaker Swapna Somasundaran

Practice talk for ACL 2006 Workshop on Frontiers in Annotation

Abstract

This paper applies the categories from an opinion annotation scheme developed for monologue text to the genre of multiparty meetings. We describe modifications to the coding guidelines that were required to extend the categories to the new type of data, and present the results of an inter-
annotator agreement study. As researchers have found with other types of
annotations in speech data, interannotator agreement is higher when the
annotators both read and listen to the data than when they only read the transcripts. Previous work exploited combinations of prosodic and lexical clues to perform automatic detection of speaker emotion (Liscombe et al. 2003). Our findings suggest that doing so to recognize opinion categories would be a promising line of work.

[TALK] Exploiting Discourse Structure for Spoken Dialogue Performance Analysis

Speaker: Mihai Rotaru

Practice talk for EMNLP 2006. Here is the paper abstract:

In this paper we study the utility of discourse structure for spoken dialogue performance modeling. We experiment with various ways of exploiting the discourse structure: in isolation, as context information for other factors (correctness and certainty) and through trajectories in the discourse structure hierarchy. Our correlation and PARADISE results show that, while the discourse structure is not useful in isolation, using the discourse structure as context information for other factors or via trajectories produces highly predictive parameters for performance analysis.