Elamur Research⌕
☰
home / papers / 2026 / PhucNguyen@DravidianLangTech 2026: Political Multiclass Sentiment Analysis with XLM-RoBERTa and Low-Rank Adaptation

PhucNguyen@DravidianLangTech 2026: Political Multiclass Sentiment Analysis with XLM-RoBERTa and Low-Rank Adaptation

Paper · 2026 · Dinh Khac Phuc Nguyen and Thìn Đặng Văn

Summary

The paper addresses seven-way political sentiment classification of code-mixed Tamil-English (Tanglish) tweets for the DravidianLangTech@ACL 2026 shared task. Informal spelling, emojis and a heavily skewed label distribution push models toward the dominant Opinionated class. The authors fine-tune XLM-RoBERTa-base with LoRA adapters rather than updating all weights. They clean the input by converting emojis into textual descriptions and removing URLs and user mentions, and they rebalance the training split through random oversampling. Learning rate, weight decay and other optimizer settings are tuned with Optuna, using dev-set macro-F1 as the objective. On the blind test set, the system reaches macro-F1 0.3763, placing second on the leaderboard and beating a full fine-tuning baseline by about 0.033. Untuned LoRA and a Focal Loss variant both score lower. The work is useful as a practical recipe for small, noisy, imbalanced datasets: systematic tuning of the training dynamics inside a parameter-efficient setup improved class balance more than changing the loss function did. Absolute scores remain modest, however.

Key points

  • Data: Tanglish tweets with 4,352 training and 544 development instances in seven classes. Opinionated makes up 31.3% of the training set and "None of the above" only 3.9%. Systems are ranked by macro-F1.

  • Model setup:

    • XLM-RoBERTa-base with LoRA on the query and value projections (r=16, α=32, dropout 0.05), training under 1% of parameters.
    • FP16 training and a maximum length of 256 tokens.
    • The best Optuna configuration used a learning rate of 1.31×10⁻⁴ with AdamW, weight decay of 0.098 and 10 epochs.

  • Test macro-F1 / accuracy:

    ConfigurationMacro-F1Accuracy
    Full fine-tuning0.34320.3033
    Standard LoRA0.31930.3033
    LoRA + Focal Loss0.32860.3511 (highest)
    LoRA + Optuna0.37630.3364

  • Error patterns: Neutral tweets are often predicted as Opinionated, and Sarcastic tweets are confused with both of those classes. The Substantiated class still retains some recall.

  • Caveat and code: The abstract mentions hashtag normalization, but the methods section describes only demojization and URL/mention removal. Code and search scripts are available at github.com/nguyenbbt/ACL2026_DravidianLangTech.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 2Research subjects and languages the work is about.Sentiment & emotion · 972 worksEfficient models & compression · 242 works

Cite

Cite as Dinh Khac Phuc Nguyen and Thìn Đặng Văn. 2026. PhucNguyen@DravidianLangTech 2026: Political Multiclass Sentiment Analysis with XLM-RoBERTa and Low-Rank Adaptation. Proceedings of the Sixth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.dravidianlangtech-1.49

@inproceedings{nguyen2026phucnguyen,
  title = {{PhucNguyen@DravidianLangTech 2026: Political Multiclass Sentiment Analysis with XLM-RoBERTa and Low-Rank Adaptation}},
  author = {Dinh Khac Phuc Nguyen and Thìn DJặng Văn},
  booktitle = {Proceedings of the Sixth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages},
  year = {2026},
  doi = {10.18653/v1/2026.dravidianlangtech-1.49},
  url = {https://doi.org/10.18653/v1/2026.dravidianlangtech-1.49}
}

References

10 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. 2019. Optuna: A next-generation hyperparameter optimization framework. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pages 2623–2631.
  2. Bharathi Raja Chakravarthi, Navya Jose, Shardul Suryawanshi, Elizabeth Sherly, and John P. McCrae. 2020. Corpus creation for sentiment analysis in code-mixed Tamil-English text. In Proceedings of the 1st Joint Workshop on Spoken Language Technologies for Under-resourced languages (SLTU) and Collabor…
  3. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116. find
  4. Adeep Hande, Ruba Priyadharshini, and Bharathi Raja Chakravarthi. 2021. Offensive language identification in low-resourced code-mixed Dravidian languages using Multilingual Transformers. In Proceedings of the Eighth Workshop on NLP for Similar Languages, Varieties and Dialects, pages 104–110. find
  5. Edward J Hu, Yelong Shen, Phil Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations.
  6. Guillaume Lemaître, Fernando Nogueira, and Christos K Aridas. 2017. Imbalanced-learn: A python toolbox to tackle the curse of imbalanced datasets in machine learning. The Journal of Machine Learning Research, 18(1):559–563. find
  7. Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. 2017. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980–2988.
  8. Ruba Priyadharshini, Bharathi Raja Chakravarthi, S. Rajalakshmi, and Premjith B. 2021. Named entity recognition for code-mixed Indian languages using Transformer models. In Proceedings of the 1st Workshop on Speech and Vision for Dravidian Languages, pages 1–7. find
  9. Mani Vegupatti, Kishore Kumar Ponnusamy, Bharathi Raja Chakravarthi, Saranya Rajiakodi, Durairaj Thenmozhi, Prasanna Kumar Kumaresan, and Sathiyaraj Thangasamy. 2026. find
  10. TamilPoliSent 2026: A Shared Task report on Multiclass Political Sentiment Analysis in Tamil. In Proceedings of the Sixth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages. Association for Computational Linguistics.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/nguyen-2026-phucnguyen-dravidianlangtech-2026-political-multiclass/