Abstract

Hate speech on social media platforms has grown to become a major problem. In this study, we explore strategies to efficiently lessen its harmful effects by supporting content moderation through machine learning (ML). In order to present a more accurate spectrum of severity and surmount the constraints of seeing hate speech as a binary task (as typical in sentiment analysis), we classify hate speech into four intensities: no hate, intimidation, offense or discrimination, and promotion of violence. For this, we first involve 31 users in annotating a dataset in English and German. To promote interpretability and transparency, we integrate our ML system in a dashboard provided with explainable AI (XAI). By performing a case study with 40 non-experts moderators, we evaluated the efficacy of the proposed XAI dashboard in supporting content moderation. Our results suggest that assessing hate intensities is important for content moderators, as these can be related to specific penalties. Similarly, XAI seems to be a promising method to improve ML trustworthiness, by this, facilitating moderators' well-informed decision-making.

Citation

BibTeX

@inproceedings{2023_ditox_hate_speech_xai,
    title = {Exploring Intensities of Hate Speech on Social Media: A Case Study on Explaining Multilingual Models with XAI},
    author = {Raisa Romanov Geleta and Klaus Eckelt and Emilia Parada-Cabaleiro and Markus Schedl},
    booktitle = {Proceedings of the 4th Conference on Language, Data and Knowledge},
    editor = {Sara Carvalho and Khan, {Anas Fahad} and {Ostroski Anic}, Ana and Blerina Spahiu and Jorge Gracia and McCrae, {John P.} and Dagmar Gromann and Barbara Heinisch and Salgado, {Ana Castro}},
    publisher = {NOVA CLUNL, Portugal},
    address = {Vienna, Austria},
    doi = {10.34619/srmk-injj},
    url = {https://aclanthology.org/2023.ldk-1.57},
    pages = {532--537},
    month = {September},
    year = {2023}
}