Culture Matters in Toxic Language Detection in Persian
Abstract: Toxic language detection is crucial for creating safer online environments and limiting the spread of harmful content. Here we show how distant supervision can expand the available datasets for Persian (Farsi) while minimizing the dependence on manual labeling. With this enriched dataset, we assess the effectiveness of various large language models (LLMs) in detecting hate speech, vulgarity, and violent content in Persian. This establishes the first comprehensive benchmark for LLMs in Persian toxic language detection. As expected, these LLMs do not perform as well on Persian toxic detection as on English. We also consider the impact of cultural context on transfer learning for toxic content detection. Specifically, we show that languages with closer cultural similarities to Persian yield better results on transfer learning. Conversely, languages with more distinct cultural differences exhibit limited improvements. This underscores the critical role of cultural alignment in enhancing the performance of transfer learning models in this domain.
Paper Type: Long
Research Area: Computational Social Science and Cultural Analytics
Research Area Keywords: Toxic Language Detection, Distant supervision, Low Resource Language, Large language models (LLMs), Transfer learning, Cross-cultural NLP
Contribution Types: Approaches to low-resource settings
Languages Studied: Persian, Farsi, Arabic, English, Indonesian
Submission Number: 1104
Loading