Question Generation for Question Answering

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 866–874 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Question Generation for Question Answering Nan Duan†, Duyu Tang†, Peng Chen§, Ming Zhou† Microsoft Research Asia† Microsoft Research and AI Group§ {nanduan,dutang,peche,mingzhou}@microsoft.com Abstract This paper presents how to generate ques- tions from given passages using neural networks, where large scale QA pairs are automatically crawled and processed from Community-QA website, and used as training data. The contribution of the pa- per is 2-fold: First, two types of question generation approaches are proposed, one is a retrieval-based method using convo- lution neural network (CNN), the other is a generation-based method using recurrent neural network (RNN); Second, we show how to leverage the generated questions to improve existing question answering sys- tems. We evaluate our question generation method for the answer sentence selection task on three benchmark datasets, includ- ing SQuAD, MS MARCO, and WikiQA. Experimental results show that, by using generated questions as an extra signal, sig- nificant QA improvement can be achieved. 1 Introduction Question Answering (or QA) is one of the core problems for AI, and consists of several typical tasks, i.e. community-based QA (Qiu and Huang, 2015), knowledge-based QA (Berant et al., 2013), text-based QA (Yu et al., 2014), and reading com- prehension (Rajpurkar et al., 2016). Most of cur- rent QA systems, e.g. (Berant and Liang, 2014), (Qiu and Huang, 2015), (Xiong et al., 2017), (Yin and Schtze, 2017), need labeled QA pairs as train- ing data. Although labeling efforts have been made, such as WebQuestions dataset (Berant et al., 2013) and SimpleQuestions dataset (Bordes et al., 2015) for knowledge-based QA, WikiQA dataset (Yang et al., 2015) for text-based QA, SQuAD dataset (Rajpurkar et al., 2016) and MS MARCO dataset (Nguyen et al., 2016) for reading compre- hension, these datasets are still with limited sizes, as labeling is very expensive. Motivated by this, we explore how to generate questions from given passages using neural net- works, with three expected goals: (1) the training data should need few or no human efforts and re- flect commonly-asked question intentions; (2) the questions are generated based on natural language passages, and should have good quality; (3) the generated questions should be helpful to QA tasks. To achieve the 1st goal, we propose to ac- quire large scale high-quality training data from Community-QA (CQA) website. The motivation of using CQA website for training data collection is that, such websites (e.g., YahooAnswers, Quo- ra, etc.) contain large scale QA pairs generated by real users, and these questions reflect the most common user intentions, and therefore are useful to search, QA, and chatbot scenarios. To achieve the 2nd goal, we explore two ways to generate questions for a given passage, one is a retrieval-based method using convolution neural network (CNN), the other is a generation-based method using recurrent neural network (RNN). We evaluate the generation quality by BLEU score (Papineni et al., 2002) and human annotations, and discuss their …

Content truncated to 500 words