Open Dialogue Management for Relational Databases

Proceedings of NAACL-HLT 2013, pages 1082–1091, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Open Dialogue Management for Relational Databases Ben Hixon Rebecca J. Passonneau Computer Science & Engineering Center for Computational Learning Systems University of Washington Columbia University Seattle, WA 98195, USA New York, New York, USA bhixon@cs.washington.edu becky@ccls.columbia.edu Abstract We present open dialogue management and its application to relational databases. An open dialogue manager generates dialogue states, ac- tions, and default strategies from the semantics of its application domain. We define three open dialogue management tasks. First, vocabulary selection finds the intelligible attributes in each database table. Second, focus discovery selects candidate dialogue foci, tables that have the most potential to address basic user goals. Third, a focus agent is instantiated for each dia- logue focus with a default dialogue strategy governed by efficiency. We demonstrate the portability of open dialogue management on three very different databases. Evaluation of our system with simulated users shows that us- ers with realistically limited domain knowledge have dialogues nearly as efficient as those of users with complete domain knowledge. 1 Introduction This paper presents open dialogue management. An open dialogue manager (ODM) generates dia- logue states, actions, and strategies from knowledge it computes about the semantics of its domain. A dialogue strategy is the procedure by which a system chooses its next action given the current state of the dialogue. The system's dialogue policy completely specifies which strategy to use in any dialogue state. Strategies can be handcrafted or learned. Reinforcement learning, the leading method for dialogue strategy learning, can yield powerful results but relies on small sets of states and actions predefined by the researcher. This reliance on domain expertise limits machine learned dialogue managers to the domains for which they were specifically designed, and contributes to the prevalence of handcrafted strategies over machine learning approaches for dialogue management in commercial applications (Paek & Pieraccini, 2008). We argue that open dia- logue management, which exploits the semantics and contents of its database to generate actions, states and default strategies, is a step towards a dialogue manager that operates across domains. As a first step to open dialogue management we present ODDMER (OPEN-DOMAIN DIALOGUE MANAGER), the first dialogue system to generate its own dialogue strategy from relational databases. ODDMER’s vocabulary selection module uses supervised learning to determine each table’s intel- ligible attributes, those most likely to be in the us- er’s vocabulary. Its focus discovery module finds candidate dialogue foci, tables that have the most potential to address basic user goals. Foci are iden- tified with schema summarization through a ran- dom walk over the database schema that ranks ta- bles by size, linguistic information, and connectivi- ty. For each candidate focus, ODDMER instanti- ates a focus agent that prompts users for values of intelligible attributes ordered by efficiency. Figure 1. ODDMER uses focus discovery and vocabu- lary selection to choose its states, actions, and strategy. 1082 This paper addresses a particular type of infor- mation-seeking dialogue in which …

Content truncated to 500 words