Bootstrapping Semantic Parsers from Conversations

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 421–432, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Bootstrapping Semantic Parsers from Conversations Yoav Artzi and Luke Zettlemoyer Computer Science & Engineering University of Washington Seattle, WA 98195 {yoav,lsz}@cs.washington.edu Abstract Conversations provide rich opportunities for interactive, continuous learning. When some- thing goes wrong, a system can ask for clari- fication, rewording, or otherwise redirect the interaction to achieve its goals. In this pa- per, we present an approach for using con- versational interactions of this type to induce semantic parsers. We demonstrate learning without any explicit annotation of the mean- ings of user utterances. Instead, we model meaning with latent variables, and introduce a loss function to measure how well potential meanings match the conversation. This loss drives the overall learning approach, which in- duces a weighted CCG grammar that could be used to automatically bootstrap the semantic analysis component in a complete dialog sys- tem. Experiments on DARPA Communica- tor conversational logs demonstrate effective learning, despite requiring no explicit mean- ing annotations. 1 Introduction Conversational interactions provide significant op- portunities for autonomous learning. A well-defined goal allows a system to engage in remediations when confused, such as asking for clarification, reword- ing, or additional explanation. The user’s response to such requests provides a strong, if often indirect, signal that can be used to learn to avoid the orig- inal confusion in future conversations. In this pa- per, we show how to use this type of conversational feedback to learn to better recover the meaning of user utterances, by inducing semantic parsers from unannotated conversational logs. We believe that this style of learning will contribute to the long term goal of building self-improving dialog systems that continually learn from their mistakes, with little or no human intervention. Many dialog systems use a semantic parsing com- ponent to analyze user utterances (e.g., Allen et al., 2007; Litman et al., 2009; Young et al., 2010). For example, in a flight booking system, the sentence Sent: I want to go to Seattle on Friday LF: λx.to(x, SEA) ∧ date(x, FRI) might be mapped to the logical form (LF) meaning representation above, a lambda-calculus expression defining the set of flights that match the user’s de- sired constraints. This LF is a representation of the semantic content that comes from the sentence, and would be input to a context-dependent understand- ing component in a full dialog system, for example to find the date that the symbol FRI refers to. To induce semantic parsers from interactions, we consider user statements in conversational logs and model their meaning with latent variables. We demonstrate that it is often possible to use the dia- log that follows a statement (including remediations such as clarifications, simplifications, etc.) to learn the meaning of the original sentence. For example, consider the first user utterance in Figure 1, where the system failed to understand the user’s request. To complete the task, the system must use a …

Content truncated to 500 words