Vocabulary Matters: A Simple yet Effective Approach to Paragraph-level Question Generation

Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 781–785 December 4 - 7, 2020. c⃝2020 Association for Computational Linguistics 781 Vocabulary Matters: A Simple yet Effective Approach to Paragraph-level Question Generation Vishwajeet Kumar IITB-Monash Research Academy Mumbai, India vishwajeet@cse.iitb.ac.in Manish Joshi IIT Bombay Mumbai, India joshimanish0511@gmail.com Ganesh Ramakrishnan IIT Bombay Mumbai, India ganesh@cse.iitb.ac.in Yuan-Fang Li Monash University Melbourne, Australia yuanfang.li@monash.edu Abstract Question generation (QG) has recently at- tracted considerable attention. Most of the current neural models take as input only one or two sentences and perform poorly when multiple sentences or complete paragraphs are given as input. However, in real-world scenar- ios, it is very important to be able to generate high-quality questions from complete para- graphs. In this paper, we present a simple yet effective technique for answer-aware question generation from paragraphs. We augment a basic sequence-to-sequence QG model with dy- namic, paragraph-specific dictionary and copy attention that is persistent across the corpus, without requiring features generated by sophis- ticated NLP pipelines or handcrafted rules. Our evaluation on SQuAD shows that our model significantly outperforms current state-of-the- art systems in question generation from para- graphs in both automatic and human evaluation. We achieve a 6-point improvement over the best system on BLEU-4, from 16.38 to 22.62. 1 Introduction and Related work Automatic question generation (QG) from text aims to generate meaningful, relevant, and answerable questions from a given textual input. Owing to its applicability in conversational systems such as Cortana, Siri, chatbots, and automated tutoring systems, QG has attracted considerable interest in both academia and industry. Recent neural network-based approaches (Du et al., 2017; Kumar et al., 2018a,b; Du and Cardie, 2018; Zhao et al., 2018; Song et al., 2018; Subramanian et al., 2018; Tang et al., 2017; Wang et al., 2017) represent the state-of-the-art in question generation. Most of these techniques learn to generate questions from short text, i.e., one or two sentences (Du et al., 2017; Kumar et al., 2018a,b; Du and Cardie, 2018). On the other hand, the ability to generate high-quality questions from longer text such as from multiple sentences or from a paragraph in its entirety, is more useful in real-world settings. However, given that a paragraph contains a longer context and more information than a sentence, it is a significantly more challenging problem to generate questions around a longer context. In figure 1 we present one motivating example demonstrating why the model needs information more than just a single sentence for generating question a meaningful and relevant question. As we can see in figure 1, question 2, question generated by our model use multiple sentences as context. Du et al. (2017) recently observed that 20% of the questions in the SQuAD dataset (Rajpurkar et al., 2016) require paragraph- level information to answer them. For the same reason, it is intuitive to conclude that the ability to consider the complete context; however long it may be, …

Content truncated to 500 words