A Semi-Supervised Approach to Detect Toxic Comments

Proceedings of Recent Advances in Natural Language Processing, pages 1261–1267 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_142 1261 A Semi-Supervised Approach to Detect Toxic Comments Ghivvago D. Saraiva1, Rafael T. Anchiˆeta2, Francisco A. R. Neto2 and Raimundo S. Moura1 1Federal University of Piau´ı, Brazil {ghivvagodamas.ufpi,rsm}@ufpi.edu.br 2Federal Institute of Piau´ı, Brazil {rta,farn}@ifpi.edu.br Abstract Toxic comments contain forms of non- acceptable language targeted towards groups or individuals. These types of comments be- come a serious concern for government orga- nizations, online communities, and social me- dia platforms. Although there are some ap- proaches to handle non-acceptable language, most of them focus on supervised learning and the English language. In this paper, we deal with toxic comment detection as a semi-supervised strategy over a heteroge- neous graph. We evaluate the approach on a toxic dataset of the Portuguese language, outperforming several graph-based methods and achieving competitive results compared to transformer architectures. 1 Introduction Toxic comments, posts, and other types of content became more common in social media nowadays. They contain forms of non-acceptable language (profanity), which may be concealed or explicit, including insults and threats directed to a group or individual (Zampieri et al., 2019). These comments spread rapidly on the internet, especially on social networks where they find acceptance, and may cul- minate in several threats to individuals, becoming a serious concern for government organizations, online communities, and social media platforms. The term toxic comment is commonly found in literature as harmful speech, hate speech, or offen- sive language. Toxic comment may be viewed as negative online behaviors, i.e., comments that are rude, disrespectful, may contain hate speech, or oth- erwise likely to make someone leave a discussion1. Schmidt and Wiegand (2017) define hate speech as any communication that disparages a person or a 1https://www.kaggle.com/c/jigsaw-toxi c-comment-classification-challenge/overv iew group based on some characteristic such as race, color, ethnicity, gender, sexual orientation, nation- ality, religion, or other characteristics. Also, it may occur with different linguistic styles, even in subtle forms or when humour is used (Fortuna and Nunes, 2018). It is important to highlight that fighting these types of comments is of utmost importance since they are a crime in several countries. To deal with toxic comments, most approaches adopt supervised-machine learning techniques and are mainly focused on the English language (Po- letto et al., 2020). These approaches range from surface-level features, as Bag-Of-Words (Paiva et al., 2019), linguistics features, as Part-Of-Speech information (Chen et al., 2012), deep neural net- works, as Long Short-Term Memory (LSTM) (For- tuna et al., 2019) and Convolutional Neural Net- works (CNN) (Badjatiya et al., 2017) to Trans- former architectures (Leite et al., 2020). Despite interesting results achieved by Transformer archi- tectures, there are still several rooms to be explored in this research area. In this paper, we developed a semi-supervised strategy to detect toxic comments in the Brazil- ian Portuguese language. Semi-supervision is the problem of learning from labeled and un- labeled data (Abney, 2007; Subramanya and Talukdar, 2014), in which given a point set X = {x1, ..., xl, xl+1, …

Content truncated to 500 words