MGR: Multi-generator Based Rationalization

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 12771–12787 July 9-14, 2023 ©2023 Association for Computational Linguistics MGR: Multi-generator Based Rationalization Wei Liu1 Haozhao Wang1∗ Jun Wang2† Ruixuan Li1 Xinyang Li1 Yuankai Zhang1 Yang Qiu1 1School of Computer Science and Technology, 1Huazhong University of Science and Technology, Wuhan City, Hubei Province, China 2iWudao Tech 1{idc_lw, hz_wang, rxli, lxy722, yuankai_zhang, anders}@hust.edu.cn 2jwang@iwudao.tech Abstract Rationalization is to employ a generator and a predictor to construct a self-explaining NLP model in which the generator selects a subset of human-intelligible pieces of the input text to the following predictor. However, rationalization suffers from two key challenges, i.e., spurious correlation and degeneration, where the predic- tor overfits the spurious or meaningless pieces solely selected by the not-yet well-trained gen- erator and in turn deteriorates the generator. Although many studies have been proposed to address the two challenges, they are usually designed separately and do not take both of them into account. In this paper, we propose a simple yet effective method named MGR to simultaneously solve the two problems. The key idea of MGR is to employ multiple gener- ators such that the occurrence stability of real pieces is improved and more meaningful pieces are delivered to the predictor. Empirically1, we show that MGR improves the F1 score by up to 20.9% as compared to state-of-the-art methods. 1 Introduction The widespread use of deep learning in NLP mod- els has led to increased concerns about interpretabil- ity. To solve this problem, Lei et al. (2016) pro- posed rationalization framework RNP in which a generator selects human-intelligible subsets (i.e., rationales) from the input text and feeds them to the subsequent predictor that maximizes the text classification accuracy, as shown in Figure 1. Un- like post-hoc approaches for explaining black-box models, the RNP framework has the built-in self- explaining ability through a cooperative game be- tween the generator and the predictor. RNP and its variants have become one of the mainstreams ∗Corresponding author † This paper is a collaborative work between the Intel- ligent and Distributed Computing Laboratory at Huazhong University of Science and Technology, and iWudao Tech. 1https://github.com/jugechengzi/Rationalization-MGR 𝑿 Generator Predictor ෠𝑌 great hotel with excellent location ! the breakfast buffet is terrific . service is wonderful . [0.1, 0.9] (Location) 𝑍 great hotel with excellent location ! the breakfast buffet is terrific . service is wonderful . 𝑌 [0, 1] 𝑚𝑖𝑛θ𝑃,θ𝐺𝐻(𝑌, ෠𝑌) Figure 1: The standard rationalization framework RNP. X,Z, ˆY ,Y represent the input text, rationale, prediction and the groundtruth label, respectively. to facilitate the interpretability of NLP models (Yu et al., 2021; Liu et al., 2022, 2023). Notably, given the versatility of the self-explaining rationalization framework, such methods have significant poten- tial for application in diverse fields such as multi- aspect recommender systems (Deng et al., 2023) and computer vision (Yuan et al., 2022). Despite its strength, rationalization schemes are notoriously hard to train. Two main training ob- stacles are the spurious correlations (Chang …

Content truncated to 500 words