KEC’S CODE CRAFTERS@DravidianLangTech 2026: Abusive Tamil Text Detection Targeting Women on Social Media

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Advanced, medium

Summary

KEC'S CODE CRAFTERS developed a classification model for the "Abusive Tamil Text Detection Targeting Women on Social Media" shared task at DravidianLangTech-2026. This model aims to automatically moderate digital toxicity on social media platforms, specifically targeting online harassment against women in linguistically diverse regions. The team trained their model on a dataset comprising 25,948 comments for training and 915 for testing, with the primary goal of classifying YouTube video content as either "Abusive" or "Non-Abusive." Addressing the complexities of the Tamil language, including its agglutinative structure and frequent code-mixing with English, the researchers designed a specialized preprocessing pipeline for denoising informal scripts. They implemented four traditional machine learning models—SVM, Logistic Regression, Random Forest, and Multinomial Naive Bayes—using TF-IDF for feature extraction. The Logistic Regression model achieved the highest performance, with an accuracy and F1 score of 0.86, following optimization of hyperparameters and decision thresholds.

Key takeaway

For NLP Engineers developing content moderation systems for regional languages like Tamil, you should prioritize designing a robust preprocessing pipeline to handle agglutinative structures and code-mixing. This research demonstrates that traditional machine learning models, specifically Logistic Regression, can achieve strong performance (0.86 F1-score) when combined with effective denoising and hyperparameter optimization. Focus your efforts on these foundational steps before exploring more complex deep learning architectures, as they offer a practical and efficient solution for abusive text detection.

Key insights

Automated moderation for regional languages is critical due to digital toxicity, especially for vulnerable groups.

Principles

• Agglutinative languages with code-mixing require specialized preprocessing for text classification.
• Traditional ML models can achieve high performance in specific abusive language detection.

Method

Design a specific preprocessing pipeline for denoising informal, code-mixed, agglutinative scripts. Extract features using TF-IDF, then apply traditional ML models, optimizing hyperparameters and decision thresholds.

In practice

• Implement specialized denoising for code-mixed, agglutinative language datasets.
• Optimize Logistic Regression for high-performance abusive text detection.

Topics

• Abusive Text Detection
• Tamil Language Processing
• Social Media Moderation
• Machine Learning Classification
• Code-Mixing

Best for: Research Scientist, AI Scientist, NLP Engineer, Data Scientist

Related on AIssential

• Multi-Stage Training for Abusive Comment Detection in Indic Languages — Multi-stage training with language preprocessing and model ensembles improves abusive comment detection while minimizing false positives.
• Building Singlish2Sinhala: A Machine Learning Approach to Sinhala Transliteration — Informal transliteration systems like Singlish present unique NLP challenges due to spelling inconsistencies.
• SV-Detect: AI-generated Text Detection with Steering Vectors — Steering vectors from frozen LLM hidden states provide a robust, interpretable signal for detecting AI-generated text.
• Anatomy of Data Repositories for the Analysis and Detection of Toxicity in Portuguese — Lexical diversity and annotation practices significantly impact cross-domain generalization in hate speech detection datasets.
• Muted: Multilingual Targeted Offensive Speech Identification and Visualization — Muted identifies targeted offensive speech using transformer models and spaCy, requiring no fine-tuning for toxic span detection.
• From Logistic Regression to GPT-2: Building a Complete Spam Detection & Sentiment Analysis Pipeline — For imbalanced text classification, prioritize F1-score and Precision-Recall AUC over accuracy and ROC-AUC.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.