Less Learn Shortcut: Analyzing and Mitigating Learning of Spurious Feature-Label Correlation

• DOI: 10.24963/ijcai.2023/560
• Publication Date: 2023-08-11T08:31:30Z
• Authors: Yanrui Du; Jing Yan; Yan Chen; Jing Liu; Sendong Zhao; Qiaoqiao She; Hua Wu; Haifeng Wang; Bing Qin
• Keywords: Spurious relationship; Feature (linguistics); Sentiment Analysis

Recent research has revealed that deep neural networks often take dataset biases as a shortcut to make decisions rather than understand tasks, leading failures in real-world applications. In this study, we focus on the spurious correlation between word features and labels models learn from biased data distribution of training data. particular, define highly co-occurring with specific label word, example containing example. Our analysis shows examples are easier for learn, while at time prediction, words significantly higher contribution models' predictions, tend assign predicted over-relying labels. To mitigate over-reliance (i.e. correlation), propose strategy Less-Learn-Shortcut (LLS): our quantifies degree down-weights them accordingly. Experimental results Question Matching, Natural Language Inference Sentiment Analysis tasks show LLS is task-agnostic can improve model performance adversarial maintaining good in-domain