Learning the Structure of Task-Driven Human-Human Dialogs

Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the ACL, pages 201–208, Sydney, July 2006. c⃝2006 Association for Computational Linguistics Learning the Structure of Task-driven Human-Human Dialogs Srinivas Bangalore AT&T Labs-Research 180 Park Ave Florham Park, NJ 07932 srini@research.att.com Giuseppe Di Fabbrizio AT&T Labs-Research 180 Park Ave Florham Park, NJ 07932 pino@research.att.com Amanda Stent Dept of Computer Science Stony Brook University Stony Brook, NY stent@cs.sunysb.edu Abstract Data-driven techniques have been used for many computational linguistics tasks. Models derived from data are generally more robust than hand-crafted systems since they better reflect the distribution of the phenomena being modeled. With the availability of large corpora of spo- ken dialog, dialog management is now reaping the benefits of data-driven tech- niques. In this paper, we compare two ap- proaches to modeling subtask structure in dialog: a chunk-based model of subdialog sequences, and a parse-based, or hierarchi- cal, model. We evaluate these models us- ing customer agent dialogs from a catalog service domain. 1 Introduction As large amounts of language data have become available, approaches to sentence-level process- ing tasks such as parsing, language modeling, named-entity detection and machine translation have become increasingly data-driven and empiri- cal. Models for these tasks can be trained to cap- ture the distributions of phenomena in the data resulting in improved robustness and adaptabil- ity. However, this trend has yet to significantly impact approaches to dialog management in dia- log systems. Dialog managers (both plan-based and call-flow based, for example (Di Fabbrizio and Lewis, 2004; Larsson et al., 1999)) have tradition- ally been hand-crafted and consequently some- what brittle and rigid. With the ability to record, store and process large numbers of human-human dialogs (e.g. from call centers), we anticipate that data-driven methods will increasingly influ- ence approaches to dialog management. A successful dialog system relies on the syn- ergistic working of several components: speech recognition (ASR), spoken language understand- ing (SLU), dialog management (DM), language generation (LG) and text-to-speech synthesis (TTS). While data-driven approaches to ASR and SLU are prevalent, such approaches to DM, LG and TTS are much less well-developed. In on- going work, we are investigating data-driven ap- proaches for building all components of spoken dialog systems. In this paper, we address one aspect of this prob- lem – inferring predictive models to structure task- oriented dialogs. We view this problem as a first step in predicting the system state of a dialog man- ager and in predicting the system utterance during an incremental execution of a dialog. In particular, we learn models for predicting dialog acts of ut- terances, and models for predicting subtask struc- tures of dialogs. We use three different dialog act tag sets for three different human-human dialog corpora. We compare a flat chunk-based model to a hierarchical parse-based model as models for predicting the task structure of dialogs. The outline of this paper is as follows: In Sec- tion 2, we review current approaches to building dialog systems. In Section 3, we review …

Content truncated to 500 words