A Bilingual Malay-English Social Media Dataset for Binary Hate Speech Detection

This dataset consists of 26,985 bilingual Malay-English social media posts curated for binary hate speech detection tasks. The data are collected and processed from five publicly available sources: HateXplain (English), HateM (Malay), Toxicity-Small (Malay), Snapshot-Twitter-2022 (Malay), and Supervised-Twitter (Malay). Among them, HateXplain, HateM, and Toxicity-Small contain manually annotated labels, while Snapshot and Supervised-Twitter are pseudo-labeled using a custom pipeline.

Each entry includes cleaned social media text, a binary hate speech label (0 = non-hate, 1 = hate), language code (en for English, ms for Malay), and the original source. The dataset is balanced across classes, with 14,642 non-hate and 12,343 hate-labeled entries. Language distribution includes 13,609 English and 13,376 Malay-language texts.

Preprocessing is conducted separately for each language. Malay-language texts underwent spelling correction, slang normalisation, placeholder standardisation (, ), and translation filtering using Malaya NLP tools and custom-built dictionary (malayslangdict.py in slang_dictionary_custom folder). English texts are processed using Ekphrasis to handle social media-specific patterns, emojis, hashtags, and slang. Low-quality or duplicate entries are removed to ensure the data quality. The dataset is stored in CSV format (UTF-8) and is suitable for training and evaluating multilingual and low-resource hate speech detection models.