SUPERNOVA@DravidianLangTech 2026: Transformer and Ensemble Approaches for Abusive Tamil Text Detection Targeting Women

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

Summary

The SUPERNOVA system, presented at DravidianLangTech@ACL 2026, addresses the critical issue of detecting abusive Tamil text targeting women on social media, a challenge due to low-resource, code-mixed, and morphologically rich language. The system explores three distinct approaches: fine-tuning the MuRIL model with class balancing and label smoothing; combining MuRIL contextual embeddings with XG-Boost and decision threshold tuning; and a lightweight ensemble utilizing character-level TF-IDF and SentenceBERT features alongside Random Forest and Extra Trees classifiers. The most effective configuration achieved an accuracy of 0.8007 and a macro F1-score of 0.7994, securing the 11th rank among participating teams. These results underscore the efficacy of multilingual transformer representations integrated with ensemble methods for this specific abusive text detection task. The project's code is publicly available.

Key takeaway

For NLP engineers developing abusive language detection systems for Dravidian languages, consider integrating multilingual transformer models like MuRIL with ensemble techniques. Your approach should include class balancing and label smoothing during fine-tuning, or combine contextual embeddings with models like XG-Boost. This strategy can yield robust performance, as demonstrated by the SUPERNOVA system's 0.8007 accuracy on Tamil text, even with code-mixed data. Explore the publicly available code for implementation details.

Key insights

Multilingual transformers combined with ensemble techniques effectively detect abusive Tamil text, even with low-resource and code-mixed data.

Principles

• Multilingual transformers excel in low-resource settings.
• Ensemble methods enhance detection performance.
• Class balancing and label smoothing improve fine-tuning.

Method

The SUPERNOVA system investigated three methods: fine-tuning MuRIL, combining MuRIL embeddings with XG-Boost, and an ensemble of TF-IDF/SentenceBERT with Random Forest/Extra Trees.

In practice

• Use MuRIL for low-resource language tasks.
• Implement ensemble models for robust classification.
• Apply class balancing for imbalanced datasets.

Topics

• Abusive Language Detection
• Tamil Language Processing
• Multilingual Transformers
• Ensemble Learning
• MuRIL
• Social Media Analysis

Code references

• Kiruthi001/SuperNova-DravidianLangTech-ACL2026

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.
• 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.
• DySem: Uncovering Dynamic Semantic Components via Multilingual Consensus for Calculating Semantic Textual Similarity — DySem uses multilingual consensus to find dynamic, sample-specific semantic dimensions for improved STS in LLMs.
• 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.
• 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.