Too Many Questions? What Can We Do? : Multiple Question Span Detection

Too Many Questions? What Can We Do? : Multiple Question Span Detection Prathyusha Danda LTRC, KCIS IIIT-Hyderabad danda.prathyusha@research.iiit.ac.in Brij Mohan Lal Srivastava INRIA France brij.srivastava@inria.fr Manish Shrivastava LTRC, KCIS IIIT-Hyderabad m.shrivastava@iiit.ac.in Abstract When a human interacts with an information retrieval chat bot, he/she can ask multiple questions at the same time. Current question answering systems can’t handle this scenario effectively. In this paper we propose an ap- proach to identify question spans in a given utterance, by posing this as a sequence label- ing problem. The model is trained and evalu- ated over 4 different freely available datasets. To get a comprehensive coverage of the com- pound question scenarios, we also synthesize a dataset based on the natural question combi- nation patterns. We exhibit improvement in the performance of the DrQA system when it encounters compound questions which sug- gests that this approach is vital for real-time human-chatbot interaction. 1 Introduction Traditional question answering systems retrieve in- formation from a knowledge-base in accordance with what is being asked in a user utterance. Ques- tions in these systems are queried in a single ques- tion format, such that there is only one question per utterance. However, most of these systems suf- fer in question-answering accuracy, especially when speakers embed multiple questions within the same utterance. QA systems like DrQA by (Chen et al., 2017) do not perform well in cases when the user utterance contains more than one question. The performance of such systems is generally subopti- mal, because the answers are generated through the assumption that exactly one question is embedded within one complete utterance. In other words, the entire utterance is processed as a single question. We propose a front end for question answering systems that detects question spans within the utterance, es- pecially when multiple questions are compounded together by the user. We report accuracies compara- ble within the utterance. In order to establish the need for such a front end, we conduct a preliminary study by first retriev- ing all the question instances in the Ubuntu dia- logue corpus. One such instance from Ubuntu di- alogue corpus is: why would you recommened arch- ?.linux ? how is it comparable to debian or ubuntu The utterance might contain more than one question based on the number of contiguous ques- tion mark clusters. Such questions exhibit com- pound question scenario. These questions are usu- ally asked to avoid setting up the context again or for brevity in the dialog. We encountered several patterns for compounding the questions. In order to obtain compound questions, we artificially syn- thesized the single question instances into relevant compound questions with the most frequent ques- tion combination patterns seen earlier. We call our dataset CompoundQA. We evaluated our Multiple Question Span Detection (MQSD) model by using it as the pre-processor to the DrQA system. We ob- serve increase in performance of the system over the compound questions data. The rest of the paper is organized as follows: Section 2 surveys the related work, Section …

Content truncated to 500 words