UTNLP at SemEval-2021 Task 5: A Comparative Analysis of Toxic Span Detection using Attention-based, Named Entity Recognition, and Ensemble Models

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 995–1002 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 995 UTNLP at SemEval-2021 Task 5: A Comparative Analysis of Toxic Span Detection using Attention-based, Named Entity Recognition, and Ensemble Models Alireza Salemi, Nazanin Sabri, Emad Kebriaei, Behnam Bahrak, Azadeh Shakery School of Electrical and Computer Engineering, College of Engineering University of Tehran, Tehran, Iran {alireza.salemi,nazanin.sabri,emad.kebriaei,bahrak,shakery}@ut.ac.ir Abstract Detecting which parts of a sentence contribute to that sentence’s toxicity—rather than provid- ing a sentence-level verdict of hatefulness— would increase the interpretability of models and allow human moderators to better under- stand the outputs of the system. This pa- per presents our team’s, UTNLP, methodol- ogy and results in the SemEval-2021 shared task 5 on toxic spans detection. We test mul- tiple models and contextual embeddings and report the best setting out of all. The exper- iments start with keyword-based models and are followed by attention-based, named entity- based, transformers-based, and ensemble mod- els. Our best approach, an ensemble model, achieves an F1 of 0.684 in the competition’s evaluation phase. 1 Introduction When social media platforms were first introduced, they allowed users to post content on any topic they wished, without restricting the type of content they were allowed to put out. This absence of restric- tions, along with the anonymity of users through these platforms (Pinsonneault and Heppel, 1997; Mondal et al., 2017), resulted in the spread of of- fensive language and hate speech online. While one might think there are only a small number of users who produce these types of hateful content, it has been shown that if social media platforms are left unmoderated, over time, the language of the community as a whole will change such that it highly correlates with the speech of hateful users rather than non-hateful ones (Mathew et al., 2020). Given the huge number of social media postings every day, manual moderation of these platforms is not a possibility. As a result, many researchers began to study automatic hate speech detection. Most studies on hate speech detection only pro- vide labels at the sentence level, showing whether the construct as a whole is toxic or not. But these types of models, offer little explanation as to why the class was predicted, making it hard for human moderators to interpret the results (Pavlopoulos et al.). In an attempt to solve the aforementioned issue, we took part in SemEval-2021 shared task 5 (Pavlopou- los et al., 2021), where we aim to detect which spans of a sentence cause it to become toxic. Our contributions are as follows: We begin our experi- mentation by evaluating a random baseline. Next, we test keyword-based methods, trying to find if toxic spans often include words that are known as hateful or negative in available word lists. We then test attention-based models, building on the hypothesis that what the attention model learns to focus on when detecting toxic speech, are the toxic spans. Afterwards, we look at the …

Content truncated to 500 words