AbuseDetect_Alchemists@DravidianLangTech 2026: A Weighted 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, Software Development & Engineering · Depth: Advanced, medium

Summary

The "AbuseDetect_Alchemists" system, presented at DravidianLangTech@ACL 2026, addresses the challenge of detecting abusive Tamil text targeting women on social media. This system frames the problem as a supervised binary classification task, categorizing social media comments as either "Abusive" or "Non-Abusive." Its processing pipeline includes a specialized preprocessing stage for emoji translation, URL removal, and entity normalization. Subsequently, it fine-tunes two distinct pre-trained transformer models, MuRIL and XLM-RoBERTa, on the task-specific data. During inference, the system employs a weighted softmax ensemble, allocating a 0.6 weight to MuRIL and 0.4 to XLM-RoBERTa. This ensemble approach yielded a Macro-F1 score of 0.8115 on the test set, surpassing the performance of each individual model. The project's code is publicly accessible on GitHub.

Key takeaway

For NLP Engineers developing content moderation systems for low-resource languages, consider implementing a weighted transformer ensemble. Your approach should include tailored preprocessing for social media text, followed by fine-tuning models like MuRIL and XLM-RoBERTa. This method can significantly improve detection accuracy, as demonstrated by the 0.8115 Macro-F1 score for abusive Tamil text, offering a robust solution for challenging linguistic contexts.

Key insights

A weighted transformer ensemble effectively detects abusive Tamil text, achieving a 0.8115 Macro-F1 score.

Principles

• Ensemble methods enhance single model performance.
• Preprocessing is crucial for social media text.
• Language-specific models improve detection.

Method

The system preprocesses Tamil social media text (emoji translation, URL removal, entity normalization), fine-tunes MuRIL and XLM-RoBERTa, then combines them via a weighted softmax ensemble (0.6 MuRIL, 0.4 XLM-RoBERTa).

In practice

• Fine-tune MuRIL and XLM-RoBERTa for Dravidian languages.
• Implement weighted softmax ensembles for classification.
• Utilize specialized preprocessing for social media data.

Topics

• Abusive Language Detection
• Tamil Language Processing
• Transformer Ensembles
• MuRIL Model
• XLM-RoBERTa
• Social Media Moderation

Code references

• meclin2345/AbuseDetect_Alchemists
• meclin2345/AbuseDetect

Best for: Research Scientist, AI Scientist, NLP Engineer

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.
• From TF-IDF to Transformers: A Comparative and Ensemble Approach to Sentiment Classification — A soft voting ensemble of diverse machine learning and transformer models significantly enhances sentiment classification accuracy, particularly for nuanced text.
• 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 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.
• Addressing Data Scarcity in Bangla Fake News Detection: An LLM-Based Dataset Augmentation Approach — LLM-based data augmentation significantly improves fake news detection in low-resource languages by generating diverse synthetic data.
• 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.

Open in AIssential →

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