Type-dependent Prompt CycleQAG : Cycle Consistency for Multi-hop Question Generation

Proceedings of the 29th International Conference on Computational Linguistics, pages 6301–6314 October 12–17, 2022. 6301 Type-dependent prompt CycleQAG : Cycle consistency for Multi-hop Question Generation Seungyeon Lee1, Minho Lee1,2,∗ 1Department of Artificial Intelligence, Kyungpook National University 2ALI Co., Ltd. statai3237@knu.ac.kr, mholee@knu.ac.kr Abstract Multi-hop question generation (QG) is the pro- cess of generating answer related questions, which requires aggregating multiple pieces of information and reasoning from different parts of the texts. This is opposed to single-hop QG which generates questions from sentences con- taining an answer in a given paragraph. Single- hop QG requires no reasoning or complexity, while multi-hop QG often requires logical rea- soning to derive an answer related question, making it a dual task. Not enough research has been made on the multi-hop QG due to its complexity. Also, a question should be created using the question type and words related to the correct answer as a prompt so that multi- hop questions can get more information. In this view, we propose a new type-dependent prompt cycleQAG (cyclic question-answer-generation), with a cycle consistency loss in which QG and Question Answering (QA) are learnt in a cyclic manner. The novelty is that the cy- cle consistency loss uses the negative cross entropy to generate syntactically diverse ques- tions that enable selecting different word repre- sentations. Empirical evaluation on the multi- hop dataset with automatic and human evalua- tion metrics outperforms the baseline model by about 10.38% based on ROUGE score. 1 Introduction Question Generation (QG) problem that automati- cally generates a question from a given document with a correct answer is a challenging and an inter- esting task in the field of natural language process- ing (Chan and Fan, 2019; Pan et al., 2021; Yu et al., 2020; Dong et al., 2019). With the advent of deep learning, the pre-trained language models (Devlin et al., 2019; Radford et al., 2018; Liu et al., 2019; Raffel et al., 2020; Clark et al., 2020; Peng et al., 2021) were proposed, after which the study of nat- ural language processing began to develop rapidly. ∗Corresponding author These works not only use single-hop QA dataset such as SQuAD (Rajpurkar et al., 2016), which is a representative of research on Question Answering (QA), but also the multi-hop QA dataset such as HotpotQA (Yang et al., 2018). The QA dataset con- sists of (Context, Question, Answer) pairs along with a lot of QA data, that enables research on Automatic Question Generation (AQG). Most of the question generation methods evaluated ques- tions using the single-hop QA datasets (Duan et al., 2017; Du et al., 2017; Sultan et al., 2020). How- ever, in real-world situations, the questions can be very complex and sometimes require a complicated reasoning process (Gupta et al., 2020; Pan et al., 2021; Yu et al., 2020). Multi-hop QG requires combining several pieces of information and reasoning over them to derive an answer related a question, making it a dual task. Multi-hop questions that can be encountered in the real world are largely divided …

Content truncated to 500 words