Explainable Question Answering based on Semantic Graph by Global Differentiable Learning and Dynamic Adaptive Reasoning

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 5318 - 5325 December 7-11, 2022 ©2022 Association for Computational Linguistics Explainable Question Answering based on Semantic Graph by Global Differentiable Learning and Dynamic Adaptive Reasoning Jianguo Mao1,2‡ Wenbin Jiang3∗ Xiangdong Wang1 Hong Liu1 Yu Xia3, Yajuan Lyu3, Qiaoqiao She3 1Beijing Key Laboratory of Mobile Computing and Pervasive Device, Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China 2 University of Chinese Academy of Sciences, Beijing, China 3 Baidu Inc., Beijing, China {maojianguo20s,xdwang,hliu}@ict.ac.cn {jiangwenbin,lvyajuan,sheqiaoqiao}@baidu.com Abstract Multi-hop Question Answering is an agent task for testing the reasoning ability. With the development of pre-trained models, the im- plicit reasoning ability has been surprisingly improved and can even surpass human perfor- mance. However, the nature of the black box hinders the construction of explainable intel- ligent systems. Several researchers have ex- plored explainable neural-symbolic reasoning methods based on question decomposition tech- niques. The undifferentiable symbolic opera- tions and the error propagation in the reasoning process lead to poor performance. To allevi- ate it, we propose a simple yet effective Global Differentiable Learning strategy to explore opti- mal reasoning paths from the latent probability space so that the model learns to solve interme- diate reasoning processes without expert anno- tations. We further design a Dynamic Adaptive Reasoner to enhance the generalization of un- seen questions. Our method achieves 17% im- provements in F1-score against BreakRC and shows better interpretability. We take a step for- ward in building interpretable reasoning meth- ods. 1 Introduction Multi-hop Question Answering involves retrieving supporting facts from multiple documents along with the explicit reasoning path and reasoning out the answer (Yang et al., 2018). As pre-trained lan- guage models evolved, the performance on this task improved spectacularly (Kenton and Toutanova, 2019; Beltagy et al., 2020; Zaheer et al., 2020; Joshi et al., 2020; Zhu et al., 2021a; Li et al., 2022). De- spite the success, the black-box nature of pure neu- ral networks has raised concerns among researchers that the unexplainable reasoning process is unac- ceptable for building trustworthy and robust intelli- gent systems (Min et al., 2019; Ding et al., 2019; *Corresponding author. ‡This work was done during internship at Baidu Inc. Multi-Hop Question: Where is a place that has both Lano Estacado and Buda Limestone geological formations? places in both #1 and #2 places that have Lano Estacado formations places that have Buda Limestone formations where are #3 Reasoner Reasoner Question Decomposition (a). Question Semantic Graph (b). BreakRC (Baseline) (c). Ours Reasoner Reasoner : Undifferentiable Operation Answer of Sub-Question #2 DA-Reasoner DA-Reasoner DA-Reasoner DA-Reasoner #1 #2 #3 #4 Select (Reasoning Type) Select Intersection Bridge Answer Slots Global Differentiable Learning Answer of Multi-Hop Question Sampling by Probability DA: Dynamic Adaptive Figure 1: Overall architecture of the proposed method. (a) gives an instance of the Question Semantic Graph. As (c) shows, we propose two simple yet effec- tive improvements for the explainable reasoning method illustrated in (b), including Global Differentiable Learn- ing and Dynamic Adaptive Reasoner …

Content truncated to 500 words