Can You Unpack That? Learning to Rewrite Questions-in-Context

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 5918–5924, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 5918 Can You Unpack That? Learning to Rewrite Questions-in-Context Ahmed Elgohary, Denis Peskov, Jordan Boyd-Graber∗ Department of Computer Science, UMIACS, iSchool, Language Science Center University of Maryland, College Park {elgohary, dpeskov, jbg}@cs.umd.edu Abstract Question answering is an AI-complete prob- lem, but existing datasets lack key elements of language understanding such as coreference and ellipsis resolution. We consider sequen- tial question answering: multiple questions are asked one-by-one in a conversation between a questioner and an answerer. Answering these questions is only possible through un- derstanding the conversation history. We in- troduce the task of question-in-context rewrit- ing: given the context of a conversation’s his- tory, rewrite a context-dependent into a self- contained question with the same answer. We construct, CANARD, a dataset of 40,527 ques- tions based on QUAC (Choi et al., 2018) and train Seq2Seq models for incorporating con- text into standalone questions. 1 Introduction Question Answering (QA) is an AI complete prob- lem (Webber, 1992), but existing QA datasets do not rise to the challenge: they lack key NLP prob- lems like anaphora resolution, coreference disam- biguation, and ellipsis resolution. The logic needed to answer these types of questions requires deeper NLP understanding that simulates the context in which humans naturally answer questions. Neural techniques question answering have im- proved (Devlin et al., 2018) machine reading com- prehension (Rajpurkar et al., 2016, MRC): comput- ers can take a single question and extract answers from datasets like Wikipedia. However, QA models struggle to generalize when questions do not look like the standalone questions systems in training data: e.g., new genres, languages, or closely-related tasks (Yogatama et al., 2019). Conversational question answering (Reddy et al., 2019, CQA) is a generalization that ask multiple ∗ Now at Google AI Zürich What happened to Anna Vissi in 1983? What happened in 1983? A1: In May 1983, she marries Nikos Karvelas, a composer Did they have any children? A2: In November, she gave birth to her daughter Sofia Did she have any other children? Question Rewriting Did Anna Vissi have any other children than her daughter Sofia? Did Anna Vissi and Nikos Karvelas have any children together? A3: I don’t know Q1: Q2: Q3: Figure 1: Question-in-context rewriting task. The in- put to each step is a question to rewrite given the dialog history which consists of the dialog utterances (ques- tions and answers) produced before the given ques- tion is asked. The output is an equivalent, context- independent paraphrase of the input question. questions in an information-seeking dialogs. Un- like MRC, CQA requires models to link questions together to resolve the conversational dependencies between them: each question needs to be under- stood in the conversation context. For example, the question “What was he like in that episode?” cannot be understood without knowing what “he” and “that episode” refer …

Content truncated to 500 words