NITC-HSR@DravidianLangTech 2026: Ensembling Multilingual Transformer Models for Detecting Abusive Tamil Text 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

A study by Rameez Mohammed A and S D Madhu Kumar for NITC-HSR@DravidianLangTech 2026 addresses the challenge of detecting misogynistic content in low-resource languages like Tamil on social media. The research investigates the effectiveness of multilingual transformer models for identifying abusive Tamil text specifically targeting women. Initial findings demonstrate that these models establish a strong baseline performance for the task. Further improvements were achieved by ensembling the two best-performing models, enhancing classification accuracy. The study also underscores the importance of domain-specific pre-training for boosting classifier performance. The top-performing ensemble model achieved a weighted F1 score of 0.83 on the test set, securing the first position in the shared task.

Key takeaway

For NLP Engineers developing abuse detection systems for low-resource languages like Tamil, you should prioritize multilingual transformer models as a strong baseline. Consider ensembling your top two models and implementing domain-specific pre-training to significantly boost performance. This approach achieved a 0.83 weighted F1 score, demonstrating its effectiveness in a competitive shared task. You can replicate these strategies to enhance your own content moderation efforts.

Key insights

Ensembling multilingual transformer models with domain-specific pre-training effectively detects abusive Tamil text.

Principles

• Multilingual transformers offer a strong baseline for low-resource language abuse detection.
• Ensembling top models enhances classification performance.
• Domain-specific pre-training improves classifier results.

Method

The approach involved investigating multilingual transformer models, then ensembling the two best-performing ones to improve classification.

In practice

• Deploy multilingual transformers for low-resource text classification.
• Combine multiple models via ensembling for performance gains.
• Prioritize domain-specific pre-training for classifier accuracy.

Topics

• Multilingual Transformers
• Abusive Text Detection
• Tamil Language Processing
• Ensemble Models
• Social Media Moderation
• Domain-Specific Pre-training

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

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.
• An End-to-End Hybrid Framework for Rumour Detection in Low-Resources Algerian Dialect — A hybrid framework combining transformer embeddings with classical classifiers enables effective rumour detection in low-resource Algerian dialect.
• mdok-style at SemEval-2026 Task 9: Finetuning LLMs for Multilingual Polarization Detection — Finetuning mid-sized LLMs with QLoRA and augmented data improves multilingual online polarization 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.
• 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.
• Detecting HIV-Related Stigma in Clinical Narratives Using Large Language Models — LLMs can effectively detect HIV-related stigma in clinical notes, with encoder models outperforming generative models in zero-shot settings.

Open in AIssential →

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