Entropy-based Attention Regularization Frees Unintended Bias Mitigation from Lists

Findings of the Association for Computational Linguistics: ACL 2022, pages 1105 - 1119 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Entropy-based Attention Regularization Frees Unintended Bias Mitigation from Lists Giuseppe Attanasio1,2, Debora Nozza1, Dirk Hovy1, Elena Baralis2 1Bocconi University, Milan, Italy 2Politecnico di Torino, Turin, Italy {giuseppe.attanasio3,debora.nozza,dirk.hovy}@unibocconi.it, elena.baralis@polito.it Abstract Warning: This paper contains examples of lan- guage that some people may find offensive. Natural Language Processing (NLP) models risk overfitting to specific terms in the training data, thereby reducing their performance, fair- ness, and generalizability. E.g., neural hate speech detection models are strongly influ- enced by identity terms like gay, or women, resulting in false positives, severe unintended bias, and lower performance. Most mitigation techniques use lists of identity terms or sam- ples from the target domain during training. However, this approach requires a-priori knowl- edge and introduces further bias if important terms are neglected. Instead, we propose a knowledge-free Entropy-based Attention Reg- ularization (EAR) to discourage overfitting to training-specific terms. An additional objec- tive function penalizes tokens with low self- attention entropy. We fine-tune BERT via EAR: the resulting model matches or exceeds state- of-the-art performance for hate speech classi- fication and bias metrics on three benchmark corpora in English and Italian. EAR also re- veals overfitting terms, i.e., terms most likely to induce bias, to help identify their effect on the model, task, and predictions. 1 Introduction Online hate speech is growing at a rapid pace, with effects that can result in dangerous criminal acts offline. Due to its verbal nature, various Natural Language Processing approaches have been pro- posed (Qian et al., 2018; Indurthi et al., 2019; At- tanasio and Pastor, 2020; Kennedy et al., 2020; Vidgen et al., 2021, inter alia). Recently, detec- tion performance has significantly improved with the use of large pre-trained language models based on Transformers (Vaswani et al., 2017), such as Bidirectional Encoder Representations from Trans- formers (BERT) (Devlin et al., 2019). However, Figure 1: False positive from BERT as a hate speech detector. The darker and taller the bar, the higher the overfitting on the term. several works have shown that by fine-tuning neu- ral language models on hate speech detection, the classifiers obtained contain severe unintended bias (Dixon et al., 2018), i.e. they perform better or worse when texts mention specific identity terms (such as gay, Muslim, or woman). As a result, a sentence like “As a Muslim woman, I agree” would be wrongly classified as hate speech, purely due to the presence of two identity terms, i.e., terms referring to specific groups based on their socio- demographic features. One cause of false positives is selection bias in the keyword-driven collection of corpora (Ousidhoum et al., 2020). Figure 1 shows a false positive example for a fine-tuned BERT model on hate speech detection. Ideally, the model should rely on the words adore and you. Instead, BERT overfitted to the word Girl and associated it with a hateful context. This unwanted effect demonstrates the issues of lexical overfitting, and how …

Content truncated to 500 words