Learning to Map Context-Dependent Sentences to Executable Formal Queries

Proceedings of NAACL-HLT 2018, pages 2238–2249 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics Learning to Map Context-Dependent Sentences to Executable Formal Queries Alane Suhr†, Srinivasan Iyer‡, and Yoav Artzi† † Dept. of Computer Science and Cornell Tech, Cornell University, New York, NY {suhr, yoav}@cs.cornell.edu ‡ Paul G. Allen School of Computer Science & Engineering, Univ. of Washington, Seattle, WA sviyer@cs.washington.edu Abstract We propose a context-dependent model to map utterances within an interaction to executable formal queries. To incorporate interaction his- tory, the model maintains an interaction-level encoder that updates after each turn, and can copy sub-sequences of previously predicted queries during generation. Our approach com- bines implicit and explicit modeling of refer- ences between utterances. We evaluate our model on the ATIS flight planning interac- tions, and demonstrate the benefits of model- ing context and explicit references. 1 Introduction The meaning of conversational utterances depends strongly on the history of the interaction. Con- sider a user querying a flight database using nat- ural language (Figure 1). Given a user utterance, the system must generate a query, execute it, and display results to the user, who then provides the next request. Key to correctly mapping utterances to executable queries is resolving references. For example, the second utterance implicitly depends on the first, and the reference ones in the third utterance explicitly refers to the response to the second utterance. Within an interactive system, this information needs to be composed with men- tions of database entries (e.g., Seattle, next Mon- day) to generate a formal executable representa- tion. In this paper, we propose encoder-decoder models that directly map user utterances to exe- cutable queries, while considering the history of the interaction, including both previous utterances and their generated queries. Reasoning about how the meaning of an utter- ance depends on the history of the interaction is critical to correctly respond to user requests. As interactions progress, users may omit previously- mentioned constraints and entities, and an increas- show me flights from seattle to boston next monday [Table with 31 flights] on american airlines [Table with 5 flights] which ones arrive at 7pm [No flights returned] show me delta flights [Table with 5 flights] . . . Figure 1: An excerpt of an interaction from the ATIS flight planning system (Hemphill et al., 1990; Dahl et al., 1994). Each request is followed by a description of the system response. ing portion of the utterance meaning must be de- rived from the interaction history. Figure 2 shows SQL queries for the utterances in Figure 1. As the interaction progresses, the majority of the gener- ated query is derived from the interaction history (underlined), rather than from the current utter- ance. A key challenge is resolving what past infor- mation is incorporated and how. For example, in the figure, the second utterance depends on the set of flights defined by the first, while adding a new constraint. The third utterance further refines this set by adding a constraint to …

Content truncated to 500 words