Acquiring Domain-Specific Dialog Information from Task-Oriented Human-Human Interaction through an Unsupervised Learning

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 955–964, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics Acquiring Domain-Specific Dialog Information from Task-Oriented Human-Human Interaction through an Unsupervised Learning Ananlada Chotimongkol Alexander I. Rudnicky Language Technologies Institute Language Technologies Institute Carnegie Mellon University Carnegie Mellon University Pittsburgh, PA 15213, USA Pittsburgh, PA 15213, USA ananlada@cs.cmu.edu air@cs.cmu.edu Abstract We describe an approach for acquiring the domain-specific dialog knowledge required to configure a task-oriented dialog system that uses human-human interaction data. The key aspects of this problem are the design of a di- alog information representation and a learning approach that supports capture of domain in- formation from in-domain dialogs. To represent a dialog for a learning purpose, we based our representation, the form-based di- alog structure representation, on an observa- ble structure. We show that this representation is sufficient for modeling phenomena that oc- cur regularly in several dissimilar task- oriented domains, including information- access and problem-solving. With the goal of ultimately reducing human annotation effort, we examine the use of unsupervised learning techniques in acquiring the components of the form-based representation (i.e. task, subtask, and concept). These techniques include statis- tical word clustering based on mutual infor- mation and Kullback-Liebler distance, TextTiling, HMM-based segmentation, and bisecting K-mean document clustering. With some modifications to make these algorithms more suitable for inferring the structure of a spoken dialog, the unsupervised learning algo- rithms show promise. 1 Introduction In recent dialog management frameworks, such as RavenClaw (Bohus and Rudnicky, 2003) and Col- lagen (Rich et al., 2001), domain-dependent com- ponents of a dialog manager are clearly separated from domain-independent components. This sepa- ration allows rapid development of a dialog man- agement module in a new task-oriented domain as dialog system developers can focus only on speci- fying domain-specific dialog information (e.g. the Dialog Task Specification in RavenClaw and Task Models in Collagen) while general dialog beha- viors (e.g. turn-taking, confirmation mechanism, and generic help) are provided by the framework. For task-oriented domains, the domain-specific dialog information is equivalent to task-specific information. Examples of the task-specific infor- mation are steps in a task and domain keywords. Specifying task-specific knowledge by hand is still a time consuming process (Feng et al., 2003). Furthermore, the hand-crafted knowledge may not reflect users’ perceptions of a task (Yankelovich, 1997). To reduce the subjectivity of system devel- opers, recorded conversations of humans perform- ing a similar task as a target dialog system have been used to help the developers design the task specification. Nevertheless, analyzing a corpus of dialogs by hand requires a great deal of human ef- fort (Bangalore et al., 2006). This paper investi- gates the feasibility of automating this dialog analysis process through a machine-learning ap- proach. By inferring the task-specific dialog in- formation automatically from human-human interaction data, the knowledge engineering effort could be reduced as the developers need to only revise learned information rather than analyzing a large amount of data. Acquiring the task-specific knowledge from a corpus of …

Content truncated to 500 words