Hate-Speech and Offensive Language Detection in Roman Urdu

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 2512–2522, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 2512 Hate-Speech and Offensive Language Detection in Roman Urdu Hammad Rizwan, Muhammad Haroon Shakeel, Asim Karim Department of Computer Science, Lahore University of Management Sciences (LUMS), Lahore, Pakistan {hammad.rizwan, m.shakeel, akarim}@lums.edu.pk Abstract The task of automatic hate-speech and offen- sive language detection in social media content is of utmost importance due to its implications in unprejudiced society concerning race, gen- der, or religion. Existing research in this area, however, is mainly focused on the English language, limiting the applicability to particu- lar demographics. Despite its prevalence, Ro- man Urdu (RU) lacks language resources, an- notated datasets, and language models for this task. In this study, we: (1) Present a lexicon of hateful words in RU, (2) Develop an anno- tated dataset called RUHSOLD consisting of 10, 012 tweets in RU with both coarse-grained and fine-grained labels of hate-speech and of- fensive language, (3) Explore the feasibility of transfer learning of five existing embedding models to RU, (4) Propose a novel deep learn- ing architecture called CNN-gram for hate- speech and offensive language detection and compare its performance with seven current baseline approaches on RUHSOLD dataset, and (5) Train domain-specific embeddings on more than 4.7 million tweets and make them publicly available. We conclude that trans- fer learning is more beneficial as compared to training embedding from scratch and that the proposed model exhibits greater robustness as compared to the baselines. 1 Introduction In the last decade, online social media platforms have become extremely popular, with users grow- ing exponentially. These platforms provide users with the freedom to express their opinions and abil- ity to interact with people of diverse groups. On one hand, this has resulted in exchanges of ideas and fostered relationships, while on the other, it is exploited to spread, incite, promote, or justify hatred, violence, and discrimination against users based on their gender, religion, race, affiliation with certain groups, and views related to certain events or subjects (e.g., politics) through hateful, offen- sive, derogatory, or obscene language. If such con- tent is left unaddressed, it has known to lead to acts of violence and conflicts on a broader scale, result- ing in problems for the protection of human rights, the rule of law, and freedom of speech, which are essentials for the development of an unprejudiced democratic society. Most of the social media platforms address this issue by employing techniques such as reporting and manual review by humans, which is limited by the reviewer’s speed, ability to understand the evolution of slang, jargon, and familiarity with mul- tilingual content. Apart from these issues, this pro- cess generally takes 24 hours, and by that time, the intended damage has already taken place. More- over, the manual process also poses problems re- lated to the subjective notions of what constitutes hate-speech and offensive language, which might result in misuse of this process to silence minori- ties and …

Content truncated to 500 words