Question Generation Using Sequence-to-Sequence Model with Semantic Role Labels

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2830–2842 May 2-6, 2023 ©2023 Association for Computational Linguistics Question Generation Using Sequence-to-Sequence Model with Semantic Role Labels Alireza Naeiji1, Aijun An1, Heidar Davoudi2, Marjan Delpisheh1, Muath Alzghool1 1Department of Electrical Engineering and Computer Science, York University, Canada 2Faculty of Science, Ontario Tech University, Canada {naeiji,aan,mdelpishe,alzghool}@yorku.ca heidar.davoudi@ontariotechu.ca Abstract Automatic generation of questions from text has gained increasing attention due to its use- ful applications. We propose a novel question generation method that combines the benefits of rule-based and neural sequence-to-sequence (Seq2Seq) models. The proposed method can automatically generate multiple questions from an input sentence covering different views of the sentence as in rule-based methods, while more complicated "rules" can be learned via the Seq2Seq model. The method utilizes semantic role labeling to convert training examples into their semantic representations, and then trains a Seq2Seq model over the semantic representa- tions. Our extensive experiments on three real- world data sets show that the proposed method significantly improves the state-of-the-art neu- ral question generation approaches. 1 Introduction Question Generation (QG) from text has gained increasing interest due to its usefulness in various applications such as educational reading compre- hension assessment (Chen et al., 2018; Kumar et al., 2018), data augmentation for training question- answering systems (Sultan et al., 2020), and re- sponse generation in conversational systems (Gu et al., 2021). We have been working on QG for the purpose of automatically creating the knowledge base (KB) for a conversational QA system of an industry partner1. The knowledge base consists of QA pairs extracted from a domain-specific docu- ment (such as car manuals). Previously, the cre- ation of their KB was done manually, which is very labor-intensive. To automate this KB generation process, we have tried both rule-based and neural sequence-to-sequence (Seq2Seq) QG methods. The rule-based methods create rules based on linguistic features that capture the relationships 1iNAGO Corporation (iNago.com) among components of a sentence and can gener- ate multiple questions from an input sentence to cover different aspects of the sentence. However, designing such rules is a very labour-intensive task as well. Also, these rules may not capture the com- plexity of ways a human asks questions (Yuan et al., 2019). As we will show in Section 4, the rule-based method does not lead to good results compared to neural Seq2Seq models. While the Seq2Seq methods achieve better re- sults, such methods are highly data-driven. For domains with limited training data (such as car manuals), relations that map the input text to ques- tions cannot be well captured. In addition, Seq2Seq models often generate a single question from an in- put text. However, multiple questions can be asked about a piece of text from different aspects. One way to generate multiple questions with Seq2Seq models from an input text is to mark the input text with different answer spans or keywords to show the focus for QG so that multiple questions may be generated from the …

Content truncated to 500 words