IRRGN: An Implicit Relational Reasoning Graph Network for Multi-turn Response Selection

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 8529 - 8541 December 7-11, 2022 ©2022 Association for Computational Linguistics IRRGN: An Implicit Relational Reasoning Graph Network for Multi-turn Response Selection Jingcheng Deng1,†, Hengwei Dai1,†, Xuewei Guo2,†, Yuanchen Ju1, Wei Peng3,4,∗ 1College of Computer and Information Science, Southwest University 2yz-intelligence Inc 3Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China 4School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China djc123234@163.com, kirobrine2000@163.com, g909336740@gmail.com juyuanchen0213@163.com, pengwei@iie.ac.cn Abstract The task of response selection in multi-turn di- alogue is to find the best option from all candi- dates. In order to improve the reasoning ability of the model, previous studies pay more atten- tion to using explicit algorithms to model the dependencies between utterances, which are deterministic, limited and inflexible. In addi- tion, few studies consider differences between the options before and after reasoning. In this paper, we propose an Implicit Relational Rea- soning Graph Network to address these issues, which consists of the Utterance Relational Rea- soner (URR) and the Option Dual Comparator (ODC). URR aims to implicitly extract depen- dencies between utterances, as well as utter- ances and options, and make reasoning with relational graph convolutional networks. ODC focuses on perceiving the difference between the options through dual comparison, which can eliminate the interference of the noise op- tions. Experimental results on two multi-turn dialogue reasoning benchmark datasets Mu- Tual and MuTualplus show that our method significantly improves the baseline of four pre- trained language models and achieves state-of- the-art performance. The model surpasses hu- man performance for the first time on the Mu- Tual dataset. Our code is released in the link.1 1 Introduction The response selection task is one of the most im- portant tasks in neural dialogue systems (Welleck et al., 2019; Demszky et al., 2020; Peng et al., 2022b; Zhang et al., 2020; Chen et al., 2021; Peng et al., 2022c,a; Zhao et al., 2022), which aims to find the most appropriate response from a set of candidate options given a historical dialogue. Most previous studies focus on matching between can- didate options and historical dialogues while ig- noring the reasoning ability of the model. This ∗Corresponding author. †Equal contribution. 1The codes are available at: https://github.com/ DJC-GO-SOLO/IRRGN Figure 1: An example from MuTual. All candidate options are semantically related to the historical dia- logue. Logical contradictions are marked with sky blue. Ground truth is marked with red. And the reasoning is red dashed lines. causes the model to choose logically incorrect or even counter-common-sense options, resulting in a poor user experience (Shum et al., 2018). On the recently released multi-turn dialogue reason- ing benchmark dataset MuTual (Cui et al., 2020), these traditional representative response selection models(Wu et al., 2017; Zhou et al., 2018) perform poorly, which indicates that comparing with match- ing, reasoning ability is more important in MuTual. Specifically, matching finds semantically related candidates, while reasoning requires obtaining log- ically consistent responses based on logical and semantic dependencies between …

Content truncated to 500 words