From Comments to Harm: A Findings Report on Abusive Tamil Text Targeting Women on Social Media Shared Task

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

The second shared task on Abusive Tamil Text Targeting Women on Social Media addressed the binary classification of abusive versus non-abusive content. This initiative released a dataset of Tamil YouTube comments and evaluated 24 submissions from 89 registered teams using macro-F1, aiming for balanced performance in a low-resource, noisy environment. Participating teams employed diverse approaches, including transformer fine-tuning, heterogeneous ensembles, classical baselines, and large language models utilizing prompting and LoRA. The top-performing system achieved a macro-F1 score of 0.8297, with many other submissions ranging between 0.79 and 0.81. Key findings indicate that transformer fine-tuning with domain-aligned encoders consistently performed well. Further performance improvements were frequently linked to Tamil-aware normalization and macro-F1-oriented calibration techniques, such as class-weighted learning and validation-based threshold tuning. These results underscore the critical role of language-aware preprocessing and precise decision calibration for effective moderation of women-targeted abusive Tamil social media text.

Key takeaway

For NLP Engineers developing moderation systems for low-resource languages like Tamil, prioritize language-aware preprocessing and fine-tuned models. You should implement domain-aligned transformer encoders and apply Tamil-specific normalization to improve performance. Additionally, calibrate your models using macro-F1-oriented techniques like class-weighted learning and validation-based threshold tuning to ensure balanced and reliable detection of abusive content. This approach is vital for robust social media moderation.

Key insights

Language-aware preprocessing and careful calibration are crucial for effective abusive text detection in low-resource settings.

Principles

• Domain-aligned encoders enhance transformer performance.
• Macro-F1 optimization requires specific calibration.
• Low-resource languages benefit from tailored normalization.

Method

The shared task involved binary classification of Tamil YouTube comments, evaluated by macro-F1. Teams used transformer fine-tuning, ensembles, classical baselines, and LLMs with prompting/LoRA.

In practice

• Fine-tune transformers with domain-specific encoders.
• Implement Tamil-aware text normalization.
• Apply class-weighted learning for balanced metrics.

Topics

• Abusive Text Detection
• Tamil NLP
• Social Media Moderation
• Transformer Fine-tuning
• Low-Resource Languages
• Macro-F1 Evaluation

Best for: Research Scientist, AI Scientist, NLP Engineer

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.
• TantraTagger: A Benchmark Dataset for Tantrayukti-Based Discourse Structure Labelling in Sanskrit Śāstra Texts — TantraTagger is a benchmark dataset for discourse structure labelling in Sanskrit Śāstra Texts.
• 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.
• 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 Comments to Trends: Studi NLP dan Computational Linguistics dalam Mendeteksi Topik Viral dari… — NLP and computational linguistics are crucial for detecting social media trends from vast, informal, and diverse comment data.
• 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.

Open in AIssential →

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