CUET_SYNTHETICA@DravidianLangTech 2026: Multi Architecture Transformer Ensemble for Detecting Abusive Tamil Text Targeting Women

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

Summary

CUET_SYNTHETICA@DravidianLangTech 2026 developed a multi-architecture transformer ensemble to detect abusive Tamil text targeting women on social media. This system addresses the challenge of identifying hostility in Tamil, a morphologically complex language often mixed with English, where indirect abuse frequently evades surface-pattern models. For the Shared Task on Abusive Tamil Text Targeting Women on Social Media DravidianLangTech@ACL 2026, the team classified Tamil YouTube comments as Abusive or Non-Abusive. Their approach involved training three distinct transformer models, each four times with varying learning rates, resulting in a total of 12 models. The final decision was made by averaging the predicted probabilities from these models. This 12-model ensemble achieved a macro F1 score of 0.8086, surpassing all individual models and securing 4th place in the shared task. The research highlights that combining Tamil-specialized and multilingual transformer models yields superior performance compared to any single-architecture method.

Key takeaway

For NLP Engineers developing abusive language detection systems for morphologically complex or code-mixed languages like Tamil, consider implementing a multi-architecture transformer ensemble. Your models will likely achieve higher performance, as demonstrated by a macro F1 of 0.8086, by averaging predictions from diverse models trained with varied learning rates. This approach, combining specialized and multilingual transformers, can overcome challenges posed by indirect hostility and linguistic complexity, improving your system's accuracy in real-world social media contexts.

Key insights

Combining diverse transformer models significantly improves abusive language detection in complex, code-mixed languages like Tamil.

Principles

• Ensemble methods enhance F1 scores.
• Multilingual models complement specialized ones.
• Indirect hostility requires robust models.

Method

Three transformer models were trained four times each with different learning rates. Their predicted probabilities were then averaged to make the final classification decision.

In practice

• Use ensemble for complex language tasks.
• Combine specialized and multilingual models.
• Experiment with diverse learning rates.

Topics

• Transformer Ensemble
• Abusive Language Detection
• Tamil NLP
• Multilingual Transformers
• Social Media Content Moderation
• Dravidian Languages

Best for: AI Scientist, NLP Engineer, Research Scientist

Related on AIssential

• Multi-Stage Training for Abusive Comment Detection in Indic Languages — A multi-stage pipeline combining language preprocessing and model ensembles effectively detects abusive Indic language comments while minimizing false positives.
• TeamHerald@CHIPSAL 2026: Hate Speech Detection and Sentiment Analysis of Nepali Memes using Transformer-based Architectures and Ensemble Learning — Ensemble strategies for meme analysis perform differently for binary vs. multi-class tasks, requiring task-specific aggregation.
• MORPHOGEN: A Multilingual Benchmark for Evaluating Gender-Aware Morphological Generation — MORPHOGEN evaluates LLM gender-aware morphological generation across French, Arabic, and Hindi, revealing performance gaps and biases.
• Comparison of Modern Multilingual Text Embedding Techniques for Hate Speech Detection Task — Supervised models with multilingual embeddings and gradient boosted trees excel at hate speech detection.
• A Comparative Study of Pretrained Transformer Models for Quranic ASR: Speech Representations, Label Formats, and Dataset Composition — Fine-tuning pretrained Transformers with specific speech representations and label formats significantly improves Quranic ASR accuracy and efficiency.
• A Multitask Transformer for Offensive Language Detection and Target Identification in HateBR — Multitask learning improves offensive language detection consistency but may reduce fine-grained target attribution.

Open in AIssential →

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