PolyTicsTamil_Alchemists@DravidianLangTech@ACL 2026: An Augmentation-Driven Focal Ensemble Model for Political Sentiment Analysis in Tamil

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

Summary

The PolyTicsTamil_Alchemists system, submitted to the DravidianLangTech@ACL 2026 shared task, addresses political multiclass sentiment analysis of Tamil X (Twitter) comments. This system classifies Tamil political tweets into seven sentiment categories, tackling challenges like severe class imbalance and semantic overlap. Its three-stage pipeline first balances the training set by augmenting minority classes using back-translation and transformer-based paraphrasing. Second, it fine-tunes XLM-RoBERTa-base with a class-weighted Focal Loss (𝛾=2) to focus learning on hard, ambiguous samples. Finally, it employs an ensemble of five models trained under Stratified 5-Fold Cross-Validation, averaging their softmax outputs during inference. The system achieved a Macro-F1 score of 0.3539 on the official test set, with its code publicly available.

Key takeaway

For Machine Learning Engineers developing sentiment analysis models for low-resource languages like Tamil or facing severe class imbalance, you should consider this three-stage approach. Implement data augmentation techniques such as back-translation and transformer-based paraphrasing to balance your training data. Fine-tune a robust model like XLM-RoBERTa-base with a class-weighted Focal Loss (𝛾=2) to prioritize difficult samples. Finally, enhance model robustness by training an ensemble of models with Stratified 5-Fold Cross-Validation and averaging their predictions.

Key insights

A three-stage pipeline combining data augmentation, class-weighted Focal Loss, and ensemble learning effectively addresses imbalanced political sentiment analysis in Tamil.

Principles

• Data augmentation can balance imbalanced datasets.
• Focal Loss improves learning on hard, ambiguous samples.
• Ensemble modeling enhances classification robustness.

Method

Augment minority classes via back-translation and transformer paraphrasing, fine-tune XLM-RoBERTa-base with Focal Loss (𝛾=2), then average softmax outputs from five Stratified 5-Fold Cross-Validation models.

In practice

• Augment minority classes using back-translation or paraphrasing.
• Employ Focal Loss (𝛾=2) for imbalanced multiclass tasks.
• Combine multiple models via softmax averaging for inference.

Topics

• Political Sentiment Analysis
• Tamil Language Processing
• Multiclass Classification
• Data Augmentation
• Focal Loss
• XLM-RoBERTa
• Ensemble Learning

Code references

• meclin2345/PolyTicsTamil_Alchemists
• meclin2345/PolyTicsTamil

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Exploring Data Augmentation and Resampling Strategies for Transformer-Based Models to Address Class Imbalance in AI Scoring of Scientific Explanations in NGSS Classroom — Targeted data augmentation significantly improves transformer-based model performance in automated scoring of imbalanced educational text data.
• 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.
• How to Fine-Tune an SLM for Emotion Recognition — Fine-tuning SLMs for fine-grained emotion recognition on imbalanced datasets requires a multi-technique approach to data balancing and loss weighting.
• Data Augmentations for Data-Constrained Language Model Pretraining — Data augmentations effectively mitigate overfitting in data-constrained language model pretraining, enabling productive multi-epoch training.
• Mitigating Data Scarcity in Psychological Defense Classification with Context-Aware Synthetic Augmentation — Context-aware synthetic data augmentation with theory-driven prompts significantly improves psychological defense mechanism classification in low-resource settings.
• Data Augmentation: A Fourier Analysis Perspective — Partial data augmentation can achieve full statistical benefits with significantly fewer samples than full group augmentation.

Open in AIssential →

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