Bridging The Gap: Entailment Fused-T5 for Open-retrieval Conversational Machine Reading Comprehension

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 15374–15386 July 9-14, 2023 ©2023 Association for Computational Linguistics Bridging The Gap: Entailment Fused-T5 for Open-retrieval Conversational Machine Reading Comprehension Xiao Zhang123, Heyan Huang123∗, Zewen Chi123, Xian-Ling Mao123 1School of Computer Science and Technology, Beijing Institute of Technology 2Beijing Engineering Research Center of High Volume Language Information Processing and Cloud Computing Applications 3Southeast Academy of Information Technology, Beijing Institute of Technology {yotta,hhy63,czw,maoxl}@bit.edu.cn Abstract Open-retrieval conversational machine read- ing comprehension (OCMRC) simulates real- life conversational interaction scenes. Ma- chines are required to make a decision of Yes/No/Inquire or generate a follow-up question when the decision is Inquire based on retrieved rule texts, user scenario, user question and dialogue history. Recent stud- ies try to reduce the information gap between decision-making and question generation, in order to improve the performance of genera- tion. However, the information gap still per- sists because these methods are still limited in pipeline framework, where decision-making and question generation are performed sepa- rately, making it hard to share the entailment reasoning used in decision-making across all stages. To tackle the above problem, we propose a novel one-stage end-to-end frame- work, called Entailment Fused-T5 (EFT), to bridge the information gap between decision- making and question generation in a global understanding manner. The extensive experi- mental results demonstrate that our proposed framework achieves new state-of-the-art per- formance on the OR-ShARC benchmark. Our model and code are publicly available1. 1 Introduction Open-retrieval conversational machine reading comprehension (OCMRC) (Gao et al., 2021) in- vestigates real-life scenes, aiming to formulate multi-turn interactions between humans and ma- chines in open-retrieval settings. As shown in Figure 1, given the user scenario and user ques- tion, machines are required to first retrieve related rule texts in the knowledge database, and then make a decision of Yes/No/Inquire or gen- erate a follow-up question when the decision is ∗Corresponding author. 1github.com/Yottaxx/EFT Figure 1: An example in the OCMRC dataset. Given the user scenario and user question, machines are required to first retrieve related rule texts in the knowledge database, and then make a decision of Yes/No/Inquire or generate a follow-up question when the decision is Inquire based on retrieved rule texts, user scenario, user question and dialogue history. Inquire based on retrieved rule texts, user sce- nario, user question and dialogue history. Previous studies (Saeidi et al., 2018; Verma et al., 2020; Lawrence et al., 2019; Zhong and Zettlemoyer, 2019; Gao et al., 2020a,b; Ouyang et al., 2021; Gao et al., 2021; Zhang et al., 2021) typically adopt pipeline frameworks based on pre- trained language models (PrLM) (Devlin et al., 2019; Clark et al., 2020; Lewis et al., 2020; Liu et al., 2020) , as shown in Figure 4, these frame- works usually consist of three stages, includ- 15374 Figure 2: The comparison between our framework and previous pipeline framework. a) Previous framework typically has three stages: entailment reasoning based decision-making, span extraction and question rephrasing. Thus, the entailment reasoning …

Content truncated to 500 words