On Generating Fact-Infused Question Variations

Proceedings of Recent Advances in Natural Language Processing, pages 335–345 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_039 335 On Generating Fact-Infused Question Variations Arthur Deschamps∗ ByteDance Ltd. Singapore arthur.deschamps1208@gmail.com Sujatha Das Gollapalli∗, See-Kiong Ng Institute of Data Science National University of Singapore {idssdg,seekiong}@nus.edu.sg Abstract To fully model human-like ability to ask questions, automatic question generation (QG) models must be able to produce multiple ex- pressions of the same question with differ- ent levels of detail. Unfortunately, existing datasets available for learning QG do not in- clude paraphrases or question variations affect- ing a model’s ability to learn this capability. We present FIRS, a dataset containing human- generated fact-infused rewrites of questions from the widely-used SQuAD dataset to ad- dress this limitation. Questions in FIRS were obtained by combining a given question with facts of entities referenced in the question. We study a double encoder-decoder model, Fact- Infused Question Generator (FIQG), for learn- ing to generate fact-infused questions from a given question. Experimental results show that FIQG effectively incorporates informa- tion from facts to add more detail to a given question. To the best of our knowledge, ours is the first study to present fact-infusion as a novel form of question paraphrasing. 1 Introduction Recently, automatic Question Generation (QG) is being addressed for generating natural language questions for a given input text passage. Viewed as the reverse of the well-studied question answering task (QA), QG has been applied in education and tutoring (Heilman and Smith, 2010; Lindberg et al., 2013), dialog systems and chatbots (Shum et al., 2018), as well as for improving QA systems (Duan et al., 2017; Tang et al., 2018). Various deep learning models are being rapidly developed for QG (Talmor and Berant, 2018; Kim et al., 2019; Tuan et al., 2020; Pan et al., 2020; Su et al., 2020; Wang et al., 2020a). However, it is only recently that QG studies have started focusing on an ∗*Equal contribution. All work was done at Institute of Data Science, NUS. important aspect of the human question generation process known as paraphrasing, or the ability to ask questions in diverse ways all expressing the same intent (Harrison and Walker, 2018; Wang et al., 2020b). Paraphrasing ability has been identified as a nec- essary aspect of learning human-like language gen- eration (Shum et al., 2018; Huang et al., 2020) and was previously studied in context of community QA (Liang et al., 2016; Kunneman et al., 2019; Hosking and Lapata, 2021). These works address the identification of synonymous and syntactic question variations such as (“What’s the weight of an elephant in kg?”; “How heavy is an ele- phant?”). In addition to synonymous variations, human beings are also adept at generating ques- tions expressing the same intent with varying level of details. For example, consider a QA pair from the SQuAD dataset1 shown in Table 1. SQuAD is one of the widely-used datasets for training QG models and contains about 100K training instances made up of an answer context, the answer string, and a “correct” question …

Content truncated to 500 words