Reducing Spurious Correlations for Answer Selection by Feature Decorrelation and Language Debiasing

Proceedings of the 29th International Conference on Computational Linguistics, pages 1753–1764 October 12–17, 2022. 1753 Reducing Spurious Correlations for Answer Selection by Feature Decorrelation and Language Debiasing Zeyi Zhong 1,2, Min Yang 1∗, Ruifeng Xu 3 1 Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences 2 University of Science and Technology of China 3 Harbin Institute of Technology (Shenzhen) {zhongzy,min.yang}@siat.ac.cn, xuruifeng@hit.edu.cn Abstract Deep neural models have become the main- stream in answer selection, yielding state-of- the-art performance. However, these models tend to rely on spurious correlations between prediction labels and input features, which in general suffer from robustness and general- ization. In this paper, we propose a novel Spurious Correlation reduction method to im- prove the robustness of the neural ANswer se- lection models (SCAN) from the sample and feature perspectives by removing the feature dependencies and language biases in answer selection. First, from the sample perspec- tive, we propose a feature decorrelation mod- ule by learning a weight for each instance at the training phase to remove the feature de- pendencies and reduce the spurious correla- tions without prior knowledge of such corre- lations. Second, from the feature perspective, we propose a feature debiasing module with contrastive learning to alleviate the negative language biases (spurious correlations) and fur- ther improve the robustness of the AS mod- els. Experimental results on three benchmark datasets show that SCAN achieves substantial improvements over strong baselines. For repro- ducibility, we will release our code and data at https://github.com/xish9/SCAN. 1 Introduction Answer selection, which aims to select the most applicable answers from an answer candidate pool, has broad applications in information retrieval (IR) and natural language processing (NLP). Con- ventional answer selection methods primarily fo- cus on designing various features, such as syntac- tic features (Li, 2003), dependency trees (Wang et al., 2007), and translation features (Surdeanu et al., 2008). However, the remarkable success of these methods relies heavily on feature engineer- ing, which is a labor-intensive and time-consuming process. *Min Yang is corresponding author. Subsequently, deep neural models (Qiu and Huang, 2015; Guo et al., 2017; Tay et al., 2017; Zhou et al., 2018) have been widely employed for answer selection and become the mainstream tech- niques for answer selection by automatically learn- ing the contextual representations of questions and answers. To capture the relationships between the question-answer pairs, different attention mecha- nisms (Zhang et al., 2017; Tay et al., 2018a; Shen et al., 2018; Yang et al., 2019a; Xie et al., 2020) have been proposed to learn the interactive fea- tures of the questions and the answers. Recently, pre-trained language models, such as BERT (De- vlin et al., 2018) and RoBERTa (Liu et al., 2019), have been proposed and applied to answer selection (Garg et al., 2020; MacAvaney et al., 2020; Zhang et al., 2021a), obtaining the state-of-the-art results. Despite the remarkable progress of previous works, these deep neural models are prone to rely on spurious correlations between input features and prediction labels, which capture the prediction cor- relations that hold …

Content truncated to 500 words