ByteBuilders@DravidianLangTech 2026: Transformer-Based Weighted Ensemble for Political Multiclass Sentiment Analysis of Tamil X (Twitter) Comments
Year: 2026
Type: conference-paper
Abstract: Our work is submitted as part of the Dravidian-LangTech 2026 shared task on Tamil political sentiment analysis.Tamil political sentiment analysis still focuses more difficulties due to linguistic diversity, informal social media expressions, sarcasm, and severe class imbalance in low-resource languages.Our paper presents a transformer based weighted ensemble approach for 7 class political sentiment classification of Tamil X (Twitter) comments.The proposed system combines XLM-RoBERTa and IndicBERT using weighted logit averaging to leverage both multilingual semantic understanding and Indicspecific linguistic representations.To address class imbalance, random oversampling, balanced class weights, and focal loss with γ = 2.0 are employed.Additionally, stratified 10-fold cross validation is used to improve robustness and it reduces overfitting.Experimental results show that the proposed ensemble model achieves an accuracy of 36.8% and a weighted F1-score of 34% on the development dataset.The results demonstrate the difficulties involved in fine grained Tamil political sentiment analysis, yet the results of ensemblebased transformer models can provide modest improvements in low-resource Tamil NLP tasks.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166897678