A Federated Approach for Hate Speech Detection

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 3248–3259 May 2-6, 2023 ©2023 Association for Computational Linguistics A Federated Approach for Hate Speech Detection Jay Gala* AI4Bharat jaygala24@gmail.com Deep Gandhi* University of Alberta drgandhi@ualberta.ca Jash Mehta* Georgia Institute of Technology jmehta73@gatech.edu Zeerak Talat MBZUAI Abstract Hate speech detection has been the subject of high research attention, due to the scale of content created on social media. In spite of the attention and the sensitive nature of the task, privacy preservation in hate speech detec- tion has remained under-studied. The major- ity of research has focused on centralised ma- chine learning infrastructures which risk leak- ing data. In this paper, we show that using federated machine learning can help address privacy the concerns that are inherent to hate speech detection while obtaining up to 6.81% improvement in terms of F1-score. 1 Introduction Content moderation is a topic that intersects across multiple fundamental rights, e.g., freedom of ex- pression and the right to privacy; and interest groups, e.g. scholars, legislators, civil society, and commercial entities (Kaye, 2019). The availability of public datasets has been crucial to the develop- ment of computational methods for hate speech detection. However, public data contains risks for those whose content is available. On the other hand, privately held data, e.g., data held by corporate enti- ties, holds risks for those who are reporting content. Such risks may be actualised through information leaks in models (Hitaj et al., 2017) or the transmis- sion of data (Shokri and Shmatikov, 2015), and can impact people’s safety and livelihood. In this work, we apply Federated Learning (FL, McMahan et al., 2017) to address the lack of pri- vacy in hate speech detection. FL is a privacy- preserving training paradigm for machine learning that jointly optimises for user privacy and model performance. We posit that privacy is necessary for users whose content is flagged and users who are flagging content alike. We thus operationalise privacy, in the context of hate speech detection *Equal contribution. Figure 1: Federated Learning: A centralised model is hosted on a server and is distributed to client devices, these compute weight updates, and transmit the updates for aggregation into the centralised model. The cen- tralised model is then redistributed to client devices. and federated learning, to mean privacy in terms of the content of reported content, and the report itself. FL is an apt training paradigm for tasks in which training data is highly sensitive, as FL is designed to mitigate risks of information leaks while also dealing with a high number of end-users, information loss, and label imbalances (Lin et al., 2022; Priyanshu and Naidu, 2021; Gandhi et al., 2022). We apply the FL algorithms FedProx (Li et al., 2020) and Adaptive Federated Optimization (FedOpt, Reddi et al., 2021) to 5 machine learning algorithms. We evaluate our approach on 8 previ- ously published datasets for hate speech detection. While using FL often implies a trade-off between privacy and performance, …

Content truncated to 500 words