Elamur Research⌕
☰
home / papers / 2026 / SigJBS@LT-EDI 2026: QLoRA-Tuned Homophobic and Transphobic Counter Narrative Generation

SigJBS@LT-EDI 2026: QLoRA-Tuned Homophobic and Transphobic Counter Narrative Generation

Paper · 2026 · Gaurangi Sinha, Rajarajeswari Palacharla and Manoj Balaji Jagadeeshan

Summary

The paper describes the SigJBS submission to Task 2 of the LT-EDI@ACL 2026 shared task. The task is to automatically write respectful counter-narratives in reply to homophobic and transphobic comments in English and Tamil. Tamil is a low-resource language here, with far less training data than English. The authors compare in-context prompting (0 to 10 examples) against supervised adaptation of a 4-bit quantized Gemma 3 12B instruction model using QLoRA. After joint bilingual training, they add a second stage that oversamples Tamil to counter the language imbalance. They also test, and reject, a variant that romanizes Tamil into SLP1. Fine-tuning proved more reliable than long few-shot prompts, which tended to produce shorter, template-like replies. Native-script Tamil also beat the transliterated input. The system placed 2nd in English and 6th in Tamil. This shows that cheap, parameter-efficient tuning works well for English, while Tamil counter-speech generation remains hard.

Key points

  • Data: The training data has 1,800 English and 800 Tamil pairs. The test sets have 66 English and 109 Tamil items.
  • Scoring: The official score averages five metrics: Distinct-2, BERTScore, PRS, QS and CCNC.
  • Model and tuning setup:
    • Backbone: unsloth/gemma-3-12b-it-unsloth-bnb-4bit.
    • LoRA settings: r=16, α=32, applied to attention and MLP projections.
    • Stage 1: 3 epochs on the combined English and Tamil data.
    • Stage 2: Tamil oversampled 3× to give 4,200 examples, trained for 1 more epoch.
    • Hardware: a single A100 40GB GPU.
  • Official results:
    • English: 86.35% overall, rank 2.
    • Tamil: 63.77% overall, rank 6.
    • The gap is largest in diversity: Distinct-2 was 25.29 for Tamil versus 69.32 for English.
  • Few-shot length effect: Moving from 0-shot to 10-shot shrank average output length. English replies fell from 53.3 to 17.9 words, and Tamil replies from 25.7 to 11.6 words.
  • Exploratory diagnostics (not official scores):
    • The SLP1 transliteration variant scored worst on all diagnostics, with a BERTScore of 70.41.
    • Tamil rebalancing raised BERTScore from 80.80 to 94.09.
    • After rebalancing, Tamil replies were longer and addressed the specific hateful claim more directly.

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

 Institutions

🎓Texas A&M UniversityCollege Station · United States26 works · 1,871 citations🎓Indian Institute of Technology KharagpurKharagpur · India24 works · 714 citations

 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

Cite

Cite as Gaurangi Sinha, Rajarajeswari Palacharla and Manoj Balaji Jagadeeshan. 2026. SigJBS@LT-EDI 2026: QLoRA-Tuned Homophobic and Transphobic Counter Narrative Generation. Proceedings of the Sixth Workshop on Language Technology for Equality, Diversity, Inclusion.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.ltedi-1.29

@inproceedings{sinha2026sigjbs,
  title = {{SigJBS@LT-EDI 2026: QLoRA-Tuned Homophobic and Transphobic Counter Narrative Generation}},
  author = {Gaurangi Sinha, Rajarajeswari Palacharla and Manoj Balaji Jagadeeshan},
  booktitle = {Proceedings of the Sixth Workshop on Language Technology for Equality, Diversity, Inclusion},
  year = {2026},
  doi = {10.18653/v1/2026.ltedi-1.29},
  url = {https://doi.org/10.18653/v1/2026.ltedi-1.29}
}

References

14 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Manoj J Balaji and HS Chinmaya. 2022. A study on sentimental analysis, homophobia-transphobia detection for dravidian languages. In CEUR Workshop Proceedings. https://ceur-ws.org, volume 3395, pages T2–7. ↗
  2. Bharathi Raja Chakravarthi. 2024. Detection of homophobia and transphobia in YouTube comments. International Journal of Data Science and Analytics, 18(1):49–68. find
  3. Yi-Ling Chung, Elizaveta Kuzmenko, Serra Sinem Tekiroglu, and Marco Guerini. 2019. CONAN -COunter NArratives through nichesourcing: a multilingual dataset of responses to fight online hate speech. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2819–…
  4. Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023. QLoRA: Efficient finetuning of quantized LLMs. arXiv preprint arXiv:2305.14314.
  5. Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. LoRA: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685.
  6. Prasanna Kumar Kumaresan, Devendra Deepak Kayande, Ruba Priyadharshini, Paul Buitelaar, and Bharathi Raja Chakravarthi. 2025. Homophobia and transphobia span identification in low-resource languages. Natural Language Processing Journal, 12:100169. find
  7. Hongyi Ling, Shubham Parashar, Sambhav Khurana, Blake Olson, Anwesha Basu, Gaurangi Sinha, Zhengzhong Tu, James Caverlee, and Shuiwang Ji. 2025. Complex llm planning via automated heuristics discovery. Preprint, arXiv:2502.19295. — [arXiv:2502.19295](https://arxiv.org/abs/2502.19295) find
  8. LT-EDI 2026 Organizers. 2026. Counter-narrative generation on homophobic and transphobic comments - LT-EDI@ACL 2026. https://www.codabench.org/competitions/11333/. Shared task overview and evaluation page, accessed March 3, 2026. ↗
  9. Sumedh Pendurkar and Guni Sharon. 2025. Policy-guided search on tree-of-thoughts for efficient problem solving with bounded language model queries. Transactions on Machine Learning Research. find
  10. Praveen Prasannan, Prasanna Kumar Kumaresan, Saranya Rajiakodi, C. N. Subalalitha, and Bharathi Raja Chakravarthi. 2025. Counter-speech generation for homophobic and transphobic social media content in Malayalam. Social Network Analysis and Mining, 15(1):87. find
  11. Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, Louis Rouillard, Thomas Mesnard, Geoffrey Cideron, Jean bastien Grill, Sabela Ramos, Edouard Yvinec, Michelle Casbon, Etienne Pot, Ivo Penche… find
  12. Serra Sinem Tekiroğlu, Yi-Ling Chung, and Marco Guerini. 2020. Generating counter narratives against online hate speech: Data and strategies. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1177–1190, Online. Association for Computational Linguistic…
  13. Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020. BERTScore: Evaluating text generation with BERT. In Proceedings of the International Conference on Learning Representations. find
  14. Michael Han Daniel Han and Unsloth team. 2023. Unsloth. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/sinha-2026-sigjbs-lt-edi-2026-qlora/