Answer-based Adversarial Training for Generating Clarification Questions

Proceedings of NAACL-HLT 2019, pages 143–155 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 143 Answer-based Adversarial Training for Generating Clarification Questions Sudha Rao∗ Microsoft Research, Redmond Sudha.Rao@microsoft.com Hal Daum´e III University of Maryland, College Park Microsoft Research, New York City me@hal3.name Abstract We present an approach for generating clar- ification questions with the goal of eliciting new information that would make the given textual context more complete. We propose that modeling hypothetical answers (to clarifi- cation questions) as latent variables can guide our approach into generating more useful clar- ification questions. We develop a Generative Adversarial Network (GAN) where the gener- ator is a sequence-to-sequence model and the discriminator is a utility function that models the value of updating the context with the an- swer to the clarification question. We eval- uate on two datasets, using both automatic metrics and human judgments of usefulness, specificity and relevance, showing that our approach outperforms both a retrieval-based model and ablations that exclude the utility model and the adversarial training. 1 Introduction A goal of natural language processing is to de- velop techniques that enable machines to process naturally occurring language. However, not all language is clear and, as humans, we may not always understand each other (Grice, 1975); in cases of gaps or mismatches in knowledge, we tend to ask questions (Graesser et al., 2008). In this work, we focus on the task of automati- cally generating clarification questions: questions that ask for information that is missing from a given linguistic context. Our clarification ques- tion generation model builds on the sequence-to- sequence approach that has proven effective for several language generation tasks (Sutskever et al., 2014; Serban et al., 2016; Yin et al., 2016; Du et al., 2017). Unfortunately, training a sequence- to-sequence model directly on (context, question) ∗This research performed when the author was still at University of Maryland, College Park. pairs yields questions that are highly generic1, cor- roborating a common finding in dialog systems (Li et al., 2016b). Our goal is to be able to generate clarification questions that are useful and specific. To achieve this, we begin with a recent observa- tion of Rao and Daum´e III (2018), who consider the task of question reranking: a good clarifica- tion question is the one whose answer has a high utility, which they define as the likelihood that this question would lead to an answer that will make the context more complete (§2.3). Inspired by this, we construct a model that first generates a question given a context, and then generates a hypotheti- cal answer to that question. Given this (context, question, answer) triple, we train a utility calcula- tor to estimate the usefulness of this question. We then show that this utility calculator can be gen- eralized using ideas for generative adversarial net- works (Goodfellow et al., 2014) for text (Yu et al., 2017), wherein the utility calculator plays the role of the “discriminator” and the question generator is the “generator” (§2.2), …

Content truncated to 500 words