On Length Divergence Bias in Textual Matching Models

Findings of the Association for Computational Linguistics: ACL 2022, pages 4187 - 4193 May 22-27, 2022 c⃝2022 Association for Computational Linguistics On Length Divergence Bias in Textual Matching Models Lan Jiang1, Tianshu Lyu2, Yankai Lin3, Meng Chong2, Xiaoyong Lyu2, Dawei Yin2 1Department of Automation, Tsinghua University 2Baidu Inc., Beijing, China 3Pattern Recognition Center, WeChat AI, Tencent Inc., China jiangl20@mails.thu.edu yankailin@tencent.com {lyutianshu, mengchong01, lvxiaoyong}@baidu.com yindawei@acm.org Abstract Despite the remarkable success deep mod- els have achieved in Textual Matching (TM) tasks, it still remains unclear whether they truly understand language or measure the se- mantic similarity of texts by exploiting statis- tical bias in datasets. In this work, we pro- vide a new perspective to study this issue — via the length divergence bias. We find the length divergence heuristic widely exists in prevalent TM datasets, providing direct cues for prediction. To determine whether TM mod- els have adopted such heuristic, we introduce an adversarial evaluation scheme which inval- idates the heuristic. In this adversarial setting, all TM models perform worse, indicating they have indeed adopted this heuristic. Through a well-designed probing experiment, we em- pirically validate that the bias of TM models can be attributed in part to extracting the text length information during training. To alle- viate the length divergence bias, we propose an adversarial training method. The results demonstrate we successfully improve the ro- bustness and generalization ability of models at the same time. 1 Introduction Textual matching is a crucial component in various NLP applications, such as information retrieval (Li and Xu, 2014), question answering (Shen and Lapata, 2007) and duplicate detection (Bilenko and Mooney, 2003). Given a pair of texts, the goal is to determine the semantic similarity between them. A lot of deep models (Chen et al., 2017; Wang et al., 2017; Pang et al., 2016; Guo et al., 2019; Wan et al., 2016) have achieved excellent performance on various TM benchmarks. However, recent work has found that current models are prone to adopting shallow heuristics in the datasets, rather than learning the underlying linguistics that they are intended to capture. This T1 - Microsoft acquires Maluuba, a startup focused on general artificial intelligence. (10) T2 - Microsoft has acquired Canadian startupMalu- uba, a company founded by University of Water- loo grads Kaheer Suleman and Sam Pasupalak that also participated in. (22) Label: Paraphrase Output: Non-paraphrase T1 - Bill would cut off aid to countries that don’t take back their illegal immigrant criminals. (15) T2 - Common Sense law faces massive opposition supposing that Aid would be cut off to countries who refuse their citizens. (19) Label: Non-paraphrase Output: Paraphrase Table 1: Examples for length divergence bias, origi- nally from Twitter-URL. "Output" is the output label by ESIM trained on the original training set. Numbers in bold are the number of words each text consists of. Model is misled by the length divergence of two texts. issue has been documented across tasks in natu- ral language understanding. In natural language arguments, for example, Niven and Kao …

Content truncated to 500 words