Incremental Parsing Models for Dialog Task Structure

Proceedings of the 12th Conference of the European Chapter of the ACL, pages 94–102, Athens, Greece, 30 March – 3 April 2009. c⃝2009 Association for Computational Linguistics Incremental Parsing Models for Dialog Task Structure Srinivas Bangalore and Amanda J. Stent AT&T Labs – Research, Inc., 180 Park Avenue, Florham Park, NJ 07932, USA {srini,stent}@research.att.com Abstract In this paper, we present an integrated model of the two central tasks of dialog management: interpreting user actions and generating system actions. We model the interpretation task as a classication prob- lem and the generation task as a predic- tion problem. These two tasks are inter- leaved in an incremental parsing-based di- alog model. We compare three alterna- tive parsing methods for this dialog model using a corpus of human-human spoken dialog from a catalog ordering domain that has been annotated for dialog acts and task/subtask information. We contrast the amount of context provided by each method and its impact on performance. 1 Introduction Corpora of spoken dialog are now widely avail- able, and frequently come with annotations for tasks/games, dialog acts, named entities and ele- ments of syntactic structure. These types of infor- mation provide rich clues for building dialog mod- els (Grosz and Sidner, 1986). Dialog models can be built ofine (for dialog mining and summariza- tion), or online (for dialog management). A dialog manager is the component of a dia- log system that is responsible for interpreting user actions in the dialog context, and for generating system actions. Needless to say, a dialog manager operates incrementally as the dialog progresses. In typical commercial dialog systems, the interpre- tation and generation processes operate indepen- dently of each other, with only a small amount of shared context. By contrast, in this paper we de- scribe a dialog model that (1) tightly integrates in- terpretation and generation, (2) makes explicit the type and amount of shared context, (3) includes the task structure of the dialog in the context, (4) can be trained from dialog data, and (5) runs in- crementally, parsing the dialog as it occurs and in- terleaving generation and interpretation. At the core of our model is a parser that in- crementally builds the dialog task structure as the dialog progresses. In this paper, we experiment with three different incremental tree-based parsing methods. We compare these methods using a cor- pus of human-human spoken dialogs in a catalog ordering domain that has been annotated for dialog acts and task/subtask information. We show that all these methods outperform a baseline method for recovering the dialog structure. The rest of this paper is structured as follows: In Section 2, we review related work. In Sec- tion 3, we present our view of the structure of task- oriented human-human dialogs. In Section 4, we present the parsing approaches included in our ex- periments. In Section 5, we describe our data and experiments. Finally, in Section 6, we present con- clusions and describe our current and future work. 2 Related Work There are two threads of …

Content truncated to 500 words