Lone Pine at SemEval-2021 Task 5: Fine-Grained Detection of Hate Speech Using BERToxic

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 967–973 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 967 Lone Pine at SemEval-2021 Task 5: Fine-Grained Detection of Hate Speech Using BERToxic Yakoob Khan, Weicheng Ma, Soroush Vosoughi Department of Computer Science, Dartmouth College {yakoob.khan.21,weicheng.ma.gr,soroush.vosoughi}@dartmouth.edu Abstract This paper describes our approach to the Toxic Spans Detection problem (SemEval-2021 Task 5). We propose BERToxic, a system that fine-tunes a pre-trained BERT model to lo- cate toxic text spans in a given text and uti- lizes additional post-processing steps to refine the boundaries. The post-processing steps in- volve (1) labeling character offsets between consecutive toxic tokens as toxic and (2) as- signing a toxic label to words that have at least one token labeled as toxic. Through experiments, we show that these two post- processing steps improve the performance of our model by 4.16% on the test set. We also studied the effects of data augmentation and ensemble modeling strategies on our system. Our system significantly outperformed the pro- vided baseline and achieved an F1-score of 0.683, placing Lone Pine in the 17th place out of 91 teams in the competition. Our code is made available at https://github.com/ Yakoob-Khan/Toxic-Spans-Detection 1 Introduction The promotion of respectful discourse has always been a core tenet of civilized societies. The Cam- bridge dictionary defines hate speech as “public speech that expresses hate or encourages violence towards a person or group based on something such as race, religion, sex, or sexual orientation.” Online platforms enable malicious actors to hide behind a cloak of anonymity and surreptitiously post toxic comments that are a “menace to democratic values, social stability and peace” (United Nations). To combat this problem, such platforms often employ human moderators to address offensive content that goes against community standards. However, mod- erators are unable to manually keep pace with the large volume of user-generated content today. This motivates the development of natural language pro- cessing systems to automatically detect hate speech and ensure that online platforms remain healthy and inclusive for all. There has been extensive research on hate speech detection, with the creation of large datasets (Wul- czyn et al., 2017) and the use of pre-trained text representations (Devlin et al., 2019) for varied modeling approaches. Competitions on offensive language identification (Zampieri et al., 2020) have further attracted attention to this topic. Prior work has hitherto focused on classification at the document-level based on various taxonomies, such as whether a given text contains offensive language or if it is targeted towards an individual or group. This line of inquiry does not identify the toxic spans that ascribe a text as hate speech. Doing so will assist human moderators to efficiently locate of- fensive content in long posts and elucidate further insight into hate speech explainability. This motivated the Toxic Spans Detection task (Pavlopoulos et al., 2021) where systems are asked to extract the list of toxic spans that attribute to a text’s toxicity. Consider the following example1: Text Because …

Content truncated to 500 words