Multilingual Hate Speech Detection in Code Mixed Social Media Text

Abstract

Social media platforms in India and other multilingual regions are dominated by mixed-language communication, where users freely combine languages such as Hindi, English, and Marathi within the same sentence. These messages are often written in Roman script, include informal spellings, slang words, and cultural expressions, and may change meaning depending on context. Because of this, identifying hate speech and offensive language in such content has become a challenging task. Many existing moderation systems are designed for single-language text and tend to misclassify code-mixed messages, either by flagging harmless expressions as harmful or by failing to detect genuinely offensive content.

This project focuses on the development of a software-based system for detecting hate speech and offensive language in multilingual, code-mixed social media text. The work is centered on Hindi–English and Marathi–English content, which is commonly seen on platforms such as Twitter, Facebook, and online discussion forums. The system follows a structured approach that includes data collection from publicly available sources, text cleaning and normalization for code-mixed language, feature extraction, and supervised learning-based classification. The goal is to classify user-generated text into meaningful categories such as hate speech, offensive language, or normal content.

The entire work is carried out using software tools only and does not involve any hardware components. Model performance is planned to be evaluated using standard measures such as accuracy, precision, recall, and F1-score. The final outcome of the project is an offline, user-friendly interface that allows basic experimentation and understanding of how automated moderation systems work for multilingual social media content. The project aims to provide a practical academic solution while highlighting the real challenges involved in multilingual hate speech detection.