On the Identification and Forecasting of Hate Speech in Inceldom

Proceedings of Recent Advances in Natural Language Processing, pages 373–384 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_042 373 On the Identification and Forecasting of Hate Speech in Inceldom Paolo Gajo, Arianna Muti, Katerina Korre, Silvia Bernardini and Alberto Barr´on-Cede˜no DIT, Universit`a di Bologna, Forl`ı, Italy paolo.gajo@studio.unibo.it {arianna.muti, aikaterini.korre, silvia.bernardini, a.barron}@unibo.it Abstract Spotting hate speech in social media posts is crucial to increase the civility of the Web and has been thoroughly explored in the NLP com- munity. For the first time, we introduce a mul- tilingual corpus for the analysis and identifica- tion of hate speech in the domain of inceldom, built from incel Web forums in English and Ital- ian, including expert annotation at the post level for two kinds of hate speech: misogyny and racism. This resource paves the way for the de- velopment of mono- and cross-lingual models for (a) the identification of hateful (misogynous and racist) posts and (b) the forecasting of the amount of hateful responses that a post is likely to trigger. Our experiments aim at improving the performance of Transformer-based models using masked language modeling pre-training and dataset merging. The results show that these strategies boost the models’ performance in all settings (binary classification, multi-label classification and forecasting), especially in the cross-lingual scenarios. Disclaimer: Due to the nature of the topic, this paper contains offensive words. 1 Introduction Hate speech can be generally defined as “language that is used to express hatred towards a targeted group or is intended to be derogatory, to humiliate, or to insult the members of the group” (Davidson et al., 2017). Detecting hate speech can be challeng- ing as there is a lack of consensus on its definition, while the use of offensive neologisms makes the task even more arduous (Fortuna et al., 2020). This is even more critical in environments frequented by incels, short for involuntary celibates, which per- tain to the so-called manosphere (Nagle, 2017, p. 75-86) and mainly comprise men unsuccessful in finding a sexual partner or significant other. Some of these individuals tend to engage in the spread of various forms of hate speech —in particular racism and misogyny— and recurrently adopt novel lexi- con in doing so (Blommaert, 2018). Such dynamic jargon causes models trained on hate speech to fail to recognize incel-specific instances of hate speech. Our contributions are the following: (i) Corpora. We introduce two unsupervised cor- pora on the inceldom domain, one in English and one in Italian. A subset of each corpus in- cludes manual annotations for different kinds of hate speech (cf. Section 3).1 The raw data can be used for domain adaptation and language modeling, among other applications. The annotation allows addressing three tasks. Binary: determine whether a post p conveys hate speech or not. Multi-label: determine whether p is misogynous and/or racist. Forecasting: Given an original post p′ (the first post in a thread), forecast the amount of hateful posts that it is likely to trigger in future responses. (ii) Masked language modeling. We perform mono- and cross-lingual …

Content truncated to 500 words