Towards Debiasing Sentence Representations

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5502–5515 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 5502 Towards Debiasing Sentence Representations Paul Pu Liang, Irene Mengze Li, Emily Zheng, Yao Chong Lim, Ruslan Salakhutdinov, Louis-Philippe Morency Machine Learning Department and Language Technologies Institute Carnegie Mellon University pliang@cs.cmu.edu Abstract As natural language processing methods are increasingly deployed in real-world scenarios such as healthcare, legal systems, and social science, it becomes necessary to recognize the role they potentially play in shaping so- cial biases and stereotypes. Previous work has revealed the presence of social biases in widely used word embeddings involving gen- der, race, religion, and other social constructs. While some methods were proposed to de- bias these word-level embeddings, there is a need to perform debiasing at the sentence-level given the recent shift towards new contextual- ized sentence representations such as ELMo and BERT. In this paper, we investigate the presence of social biases in sentence-level rep- resentations and propose a new method, SENT- DEBIAS, to reduce these biases. We show that SENT-DEBIAS is effective in removing biases, and at the same time, preserves per- formance on sentence-level downstream tasks such as sentiment analysis, linguistic accept- ability, and natural language understanding. We hope that our work will inspire future re- search on characterizing and removing social biases from widely adopted sentence represen- tations for fairer NLP. 1 Introduction Machine learning tools for learning from language are increasingly deployed in real-world scenarios such as healthcare (Velupillai et al., 2018), legal systems (Dale, 2019), and computational social sci- ence (Bamman et al., 2016). Key to the success of these models are powerful embedding layers which learn continuous representations of input in- formation such as words, sentences, and documents from large amounts of data (Devlin et al., 2019; Mikolov et al., 2013). Although word-level em- beddings (Pennington et al., 2014; Mikolov et al., 2013) are highly informative features useful for a variety of tasks in Natural Language Process- ing (NLP), recent work has shown that word-level embeddings reflect and propagate social biases present in training corpora (Lauscher and Glavaˇs, 2019; Caliskan et al., 2017; Swinger et al., 2019; Bolukbasi et al., 2016). Machine learning systems that incorporate these word embeddings can further amplify biases (Sun et al., 2019b; Zhao et al., 2017; Barocas and Selbst, 2016) and unfairly discrimi- nate against users, particularly those from disad- vantaged social groups. Fortunately, researchers working on fairness and ethics in NLP have devised methods towards debiasing these word representa- tions for both binary (Bolukbasi et al., 2016) and multiclass (Manzini et al., 2019) bias attributes such as gender, race, and religion. More recently, sentence-level representations such as ELMo (Peters et al., 2018), BERT (De- vlin et al., 2019), and GPT (Radford et al., 2019) have become the preferred choice for text sequence encoding. When compared to word-level represen- tations, these models have achieved better perfor- mance on multiple tasks in NLP (Wu and Dredze, 2019), multimodal learning (Zellers et …

Content truncated to 500 words