Alignment Rationale for Natural Language Inference

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 5372–5387 August 1–6, 2021. ©2021 Association for Computational Linguistics 5372 Alignment Rationale for Natural Language Inference Zhongtao Jiang1,2, Yuanzhe Zhang1,2, Zhao Yang1,2, Jun Zhao1,2 and Kang Liu1,2 1National Laboratory of Pattern Recognition, Institute of Automation, CAS, Beijing, China 2School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China {zhongtao.jiang, yzzhang, zhao.yang, jzhao, kliu}@nlpr.ia.ac.cn Abstract Deep learning models have achieved great suc- cess on the task of Natural Language Infer- ence (NLI), though only a few attempts try to explain their behaviors. Existing explana- tion methods usually pick prominent features such as words or phrases from the input text. However, for NLI, alignments among words or phrases are more enlightening clues to explain the model. To this end, this paper presents AREC, a post-hoc approach to generate align- ment rationale explanations for co-attention based models in NLI. The explanation is based on feature selection, which keeps few but suf- ficient alignments while maintaining the same prediction of the target model. Experimental results show that our method is more faith- ful and readable compared with many exist- ing approaches. We further study and re- evaluate three typical models through our ex- planation beyond accuracy, and propose a sim- ple method that greatly improves the model ro- bustness.1 1 Introduction Natural Language Inference (NLI) is a fundamental task in Natural Language Processing (NLP) which is to determine if a hypothesis entails a premise. Recently, with the introduction of large-scale an- notated datasets (Bowman et al., 2015; Williams et al., 2018), deep learning models are adopted to solve the task in a supervised manner (Conneau et al., 2017; Chen et al., 2017; Devlin et al., 2019) and achieve great success, while inner mechanisms of these methods are still opaque due to high com- putational complexities. Towards interpretability, explaining the model behavior has gained increasing attention. Lots of approaches are based on feature attribution which 1Our code is available at https://github.com/ changmenseng/arec (c) Alignment Attribution (Gradient) (a) Text Attribution (LIME) p: Passengers in a rusty yellow car driving down the street . h: People walk through a store . Passengers in a rusty yellow car driving down the street . (b) Text Rationale (Erasure) p: Passengers in a rusty yellow car driving down the street . h: People walk through a store . People walk through a store . People walk through a store . (d) Alignment Rationale (AREC) Figure 1: Different post-hoc explanations. For attribu- tion explanations, features with deeper colors are con- sidered more important. assigns saliency scores for input features (Bah- danau et al., 2015; Lundberg and Lee, 2017; Thorne et al., 2019; Kim et al., 2020), and feature selection or rationale that keeps a subset of features suffi- cient for the prediction (Lei et al., 2016; Bastings et al., 2019; De Cao et al., 2020; DeYoung et al., 2020). Figure 1 (a) and (b) present a …

Content truncated to 500 words