Non-sentential Question Resolution using Sequence to Sequence Learning

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2022–2031, Osaka, Japan, December 11-17 2016. Non-sentential Question Resolution using Sequence to Sequence Learning Vineet Kumar IBM Research Labs New Delhi, India vineeku6@in.ibm.com Sachindra Joshi IBM Research Labs New Delhi, India jasachind@in.ibm.com Abstract An interactive Question Answering (QA) system frequently encounters non-sentential (incom- plete) questions. These non-sentential questions may not make sense to the system when a user asks them without the context of conversation. The system thus needs to take into account the conversation context to process the incomplete question. In this work, we present a recurrent neural network (RNN) based encoder decoder network that can generate a complete (intended) question, given an incomplete question and conversation context. RNN encoder decoder net- works have been show to work well when trained on a parallel corpus with millions of sentences, however it is extremely hard to obtain conversation data of this magnitude. We therefore propose to decompose the original problem into two separate simplified problems where each problem focuses on an abstraction. Specifically, we train a semantic sequence model to learn semantic patterns, and a syntactic sequence model to learn linguistic patterns. We further combine syntac- tic and semantic sequence models to generate an ensemble model. Our model achieves a BLEU score of 30.15 as compared to 18.54 using a standard RNN encoder decoder model. 1 Introduction Question Answering (QA) systems (Green Jr et al., 1961; Winograd, 1971; Woods and Kaplan, 1977; Hickl et al., 2006; Gobeill et al., 2009) enable a user to obtain precise information. A natural extension is an interactive and dialogue based QA system that allows a user to ask follow up or related questions. Interactive QA system however comes with its unique set of challenges. Users ask a follow up or re- lated question by being as terse as possible, and they implicitly refer to concepts and entities in the past conversation. Table 1 depicts a few instances of follow up questions users may ask in an ongoing conversation. Incomplete questions are a subset of non-sentential utterances (NSU) (Fern´andez, 2006). NSUs are incomplete utterances which make complete sense when seen in conjunction with the utterances in con- versation. Table 1 illustrates some examples of NSU questions (Q2) a user might ask the system given a previous question (Q1) and an answer (A1). R1 refers to the intended complete question. Note that (a) and (c) need the previous question Q1, (b) needs previous answer A1, whereas (d) needs both Q1 and A1 to generate R1. The system thus either needs to restrict how users interact (Carbonell, 1983), or needs to handle the NSU questions by considering the conversation context. Restricting how users interact with a QA system is not natural, and thus can make the system hard to use. In this work, we focus on using the incomplete question and the conversation context to generate the resolved (intended) question. In the rest of the paper, we refer to this problem as NSU …

Content truncated to 500 words