Generating Highly Relevant Questions

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 5983–5987, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 5983 Generating Highly Relevant Questions Jiazuo Qiu and Deyi Xiong∗ School of Computer Science and Technology, Soochow University, Suzhou, China qjzhzw@163.com; dyxiong@suda.edu.cn Abstract The neural seq2seq based question generation (QG) is prone to generating generic and undi- versified questions that are poorly relevant to the given passage and target answer. In this paper, we propose two methods to address the issue. (1) By a partial copy mechanism, we prioritize words that are morphologically close to words in the input passage when generating questions; (2) By a QA-based reranker, from the n-best list of question candidates, we se- lect questions that are preferred by both the QA and QG model. Experiments and analyses demonstrate that the proposed two methods substantially improve the relevance of gener- ated questions to passages and answers. 1 Introduction Question generation is to generate a valid and fluent question according to a given passage and the target answer. In general, the answer is a span of words in the passage. QG can be used in many scenarios, such as automatical tutoring systems, improving the performance of QA models and en- abling chatbots to lead a conversation. In the early days, QG has been tackled mainly via rule-based approaches (Mitkov and Ha, 2003; Heilman and Smith, 2010). These methods rely on many handcrafted rules and templates. Construct- ing such rules is time-consuming and it is difficult to adapt them to other domains. Very recently, neural networks have been used for QG. Specifically, the encoder-decoder seq2seq model (Du et al., 2017; Zhou et al., 2017; Song et al., 2018) is used to encode a passage and gen- erate a question corresponding to the answer. Such end-to-end neural models are able to gen- erate better questions than traditional rule-based approaches. However, one issue with the current ∗Corresponding author neural models is that the generated questions are not quite relevant to the corresponding passages and target answers. In other words, the neural QG models tend to generate generic questions (e.g., “what is the name of...?”). In this paper, we propose two methods to deal with this low-relevance issue for QG. First, we present a partial copy method to enhance the exist- ing copy mechanism so that the QG model can not only copy a word as an entire unit from the passage to the generated question but also copy a part of a word (e.g., “start” in “started”) to generate a new morphological form of the word in the question. The fine-grained partial copy mechanism enables the QG model to copy morphologically changed words from the passage, increasing the relevance of the generated question to the passage by sharing more words in different forms. Second, we pro- pose a QA-based reranker to rerank QG results. Particularly, we use a neural QA model to evalu- ate the …

Content truncated to 500 words