How Far Can It Go? On Intrinsic Gender Bias Mitigation for Text Classification

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 3418–3433 May 2-6, 2023 ©2023 Association for Computational Linguistics How Far Can It Go? On Intrinsic Gender Bias Mitigation for Text Classification Ewoenam Kwaku Tokpo1, Pieter Delobelle2,3,5, Bettina Berendt2,3,4,5 and Toon Calders1 1 Department of Computer Science, University of Antwerp, Belgium 2 Department of Computer Science, KU Leuven, Belgium 3 Leuven.AI Institute for Artificial Intelligence, Belgium 4 Faculty of Electrical Engineering and Computer Science, TU Berlin, Germany 5 Weizenbaum Institute, Germany Abstract To mitigate gender bias in contextualized lan- guage models, different intrinsic mitigation strategies have been proposed, alongside many bias metrics. Considering that the end use of these language models is for downstream tasks like text classification, it is important to under- stand how these intrinsic bias mitigation strate- gies actually translate to fairness in downstream tasks and the extent of this. In this work, we de- sign a probe to investigate the effects that some of the major intrinsic gender bias mitigation strategies have on downstream text classifica- tion tasks. We discover that instead of resolving gender bias, intrinsic mitigation techniques and metrics are able to hide it in such a way that significant gender information is retained in the embeddings. Furthermore, we show that each mitigation technique is able to hide the bias from some of the intrinsic bias measures but not all, and each intrinsic bias measure can be fooled by some mitigation technique, but not all. We confirm experimentally, that none of the intrinsic mitigation techniques used without any other fairness intervention is able to con- sistently impact extrinsic bias. We recommend that intrinsic bias mitigation techniques should be combined with other fairness interventions for downstream tasks. 1 Introduction The use of pretrained language models has seen a surge in popularity as a result of state-of-the-art per- formances that have been achieved with these mod- els on various tasks. Consequently, there has been a growing interest in exploring how gender bias per- tains in these models (Garrido-Muñoz et al., 2021). Pretrained language models are used in two distinct phases: the pretraining phase and the finetuning phase. The pretraining phase typically involves training a model on a generic task such as masked language modeling on a diverse set of text corpora. In the finetuning phase, the pretrained model can be adapted for a specific task, such as sentiment analysis, by training on a domain-specific corpus. Owing to the unique way of using pretrained mod- els, bias generally manifests in two forms: intrinsic bias and extrinsic bias. Intrinsic bias refers to bias that inherently exists in pretrained language mod- els whereas extrinsic bias is used to refer to bias that exists in downstream models that are based on the pretrained model. Since the success of down- stream NLP tasks has mostly been dependent on pretrained models, it is intuitive to assume that bias in intrinsic models will compromise fairness in downstream tasks. Only recently have more in- depth examinations been done …

Content truncated to 500 words