--- title: "CUET-823 at SemEval-2026 Task 9: LoRA-Based Instruction" description: "The authors tackle Subtask 1 of SemEval-2026 Task 9 (POLAR) for Bengali. 9 references, cited by 1 paper here." authors: ["Arpita Mallik", "Ratnajit Dhar"] year: 2026 venue: "Proceedings of the 20th International Workshop on Semantic Evaluation (2026)" source: "https://aclanthology.org/2026.semeval-1.154/" doi: "10.18653/v1/2026.semeval-1.154" licence: "CC BY 4.0 (ACL Anthology)" licence_url: "https://creativecommons.org/licenses/by/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/mallik-2026-cuet-823-semeval-2026-task/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/mallik-2026-cuet-823-semeval-2026-task/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/11/) / CUET-823 at SemEval-2026 Task 9: LoRA-Based Instruction Fine-Tuning of LLMs vs. Transformer Models for Bengali Polarization Detection # CUET-823 at SemEval-2026 Task 9: LoRA-Based Instruction Fine-Tuning of LLMs vs. Transformer Models for Bengali Polarization Detection **Authors:** [Arpita Mallik](https://md.research.elamur.ai/free/en/authors/arpita-mallik/) and [Ratnajit Dhar](https://md.research.elamur.ai/free/en/authors/ratnajit-dhar/) · **Year:** 2026 · **Venue:** [Proceedings of the 20th International Workshop on Semantic Evaluation (2026)](https://md.research.elamur.ai/free/en/venues/semeval/) ([SemEval 2026](https://md.research.elamur.ai/free/en/venues/semeval/2026/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://aclanthology.org/2026.semeval-1.154/) · [PDF](https://aclanthology.org/2026.semeval-1.154.pdf) · [DOI 10.18653/v1/2026.semeval-1.154](https://doi.org/10.18653/v1/2026.semeval-1.154) · [OpenAlex](https://openalex.org/W7166847262) **Citations:** 0 citations (OpenAlex, 2026-10-14) · [1](https://md.research.elamur.ai/free/en/papers/mallik-2026-cuet-823-semeval-2026-task/cited-by/) in our library, of the 149,439 papers whose bibliographies we read ## Summary The authors tackle Subtask 1 of SemEval-2026 Task 9 (POLAR) for Bengali. The goal is to label a social media post as polarized or not. Here polarization covers "us versus them" rhetoric: stereotyping, vilification, dehumanization and intolerance toward other groups or views. They use the official Bengali dataset of 5,000 short posts. A lexical analysis shows that much of the test vocabulary is missing from training, so models must generalize from context rather than memorize words. Two families of systems are compared: - encoder models (mBERT, XLM-RoBERTa, MuRIL, Bangla-BERT) fine-tuned with a classification head, plus an ensemble of all four; - open LLMs adapted through 4-bit LoRA instruction fine-tuning, each tested with Bengali and English prompts. Qwen3-14B with a Bengali prompt performed best, reaching macro-F1 0.8582 and 2nd place out of 49 teams. Error analysis shows it is more sensitive than MuRIL to implicit and metaphorical hostility. The study suggests that parameter-efficient LLM adaptation is a strong option for low-resource polarization detection. It also finds that the effect of prompt language depends on the model. ## Key points - **Dataset:** 3,333 train / 166 dev / 1,501 test posts, about 42% polarized. The rare-token ratio rises from 0.11 (train) to 0.17 (test), and roughly 45% of test token types are unseen in training. - **Encoders:** MuRIL was the best single encoder (macro-F1 0.8400). The four-model ensemble added little (0.8407), and Bangla-BERT trailed the multilingual models (0.8031). - **LLM setup:** fine-tuning via Unsloth, LoRA r=32, α=32, applied to attention and MLP projections. Prediction is a single greedy-decoded token (0 or 1). - **LLM results:** Qwen3-14B scored 0.8582 with the Bengali prompt and 0.8541 with the English one. LLaMA-8B did slightly better with English (0.8493 vs. 0.8473). Mistral-7B collapsed to predicting only the majority class (0.3664). - **Qwen3-14B vs. MuRIL:** false negatives fell from 143 to 91, while false positives rose slightly from 113 to 124. Both models struggled with hostile posts that lack a clear target group. *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 - [Chittagong University of Engineering & Technology](https://md.research.elamur.ai/free/en/universities/chittagong-university-of-engineering-technology/) — University, Chittagong, Bangladesh · 25 works · 29 citations 📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works ## Cite Cite as: Arpita Mallik and Ratnajit Dhar. 2026. CUET-823 at SemEval-2026 Task 9: LoRA-Based Instruction Fine-Tuning of LLMs vs. Transformer Models for Bengali Polarization Detection. Proceedings of the 20th International Workshop on Semantic Evaluation (2026). The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{mallik2026cuet, title = {{CUET-823 at SemEval-2026 Task 9: LoRA-Based Instruction Fine-Tuning of LLMs vs. Transformer Models for Bengali Polarization Detection}}, author = {Arpita Mallik and Ratnajit Dhar}, booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)}, year = {2026}, doi = {10.18653/v1/2026.semeval-1.154}, url = {https://doi.org/10.18653/v1/2026.semeval-1.154} } ``` - [← previous CUCLASIC at SemEval-2026 Task 5: LLM Prompting Strategies for Rating Ambiguous Word Senses](https://md.research.elamur.ai/free/en/papers/riba-2026-cuclasic-semeval-2026-task-5/) - [next → CUET_InferX@DravidianLangTech 2026: Shared Task on Dialect Based Speech Recognition and Classification in Tamil](https://md.research.elamur.ai/free/en/papers/semon-2026-cuet-inferx-dravidianlangtech-2026-shared/) - [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/11/) ## References 9 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Shad Al Kaiser, Sudipta Mandal, Ashraful Kalam Abid, Ekhfa Hossain, Ferdous Bin Ali, and Intisar Tahmid Naheen. 2021. Social media opinion mining based on bangla public post of facebook. In 2021 24th International Conference on Computer and Information Technology (ICCIT), pages 1–6. IEEE. [find](https://scholar.google.com/scholar?q=Shad+Al+Kaiser%2C+Sudipta+Mandal%2C+Ashraful+Kalam+Abid%2C+Ekhfa+Hossain%2C+Ferdous+Bin+Ali%2C+and+Intisar+Tahmid+Naheen%2E+2021%2E+Social+media+opinion+mining+based+on+bangla+public+post+of+facebook%2E+In+2021+24th+) 2. Mahmud Hasan, Labiba Islam, Ismat Jahan, Sabrina Mannan Meem, and Rashedur M Rahman. 2023. Natural language processing and sentiment analysis on bangla social media comments on russia–ukraine war using transformers. Vietnam Journal of Computer Science, 10(03):329–356. [find](https://scholar.google.com/scholar?q=Mahmud+Hasan%2C+Labiba+Islam%2C+Ismat+Jahan%2C+Sabrina+Mannan+Meem%2C+and+Rashedur+M+Rahman%2E+2023%2E+Natural+language+processing+and+sentiment+analysis+on+bangla+social+media+comments+on+russia%E2%80%93ukraine+war+usin) 3. Khondoker Ittehadul Islam, Sudipta Kar, Md Saiful Islam, and Mohammad Ruhul Amin. 2021. Sentnob: A dataset for analysing sentiment on noisy bangla texts. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 3265–3271. [find](https://scholar.google.com/scholar?q=Khondoker+Ittehadul+Islam%2C+Sudipta+Kar%2C+Md+Saiful+Islam%2C+and+Mohammad+Ruhul+Amin%2E+2021%2E+Sentnob%3A+A+dataset+for+analysing+sentiment+on+noisy+bangla+texts%2E+In+Findings+of+the+Association+for+Computation) 4. Nusrat Jahan Lia, Shubhashis Roy Dipta, Abdullah Khan Zehady, Naymul Islam, Madhusodan Chakraborty, and Abdullah Al Wasif. 2025. Read between the lines: A benchmark for uncovering political bias in bangla news articles. In Proceedings of the Second Workshop on Bangla Language Processing (BLP-2025),… [find](https://scholar.google.com/scholar?q=Nusrat+Jahan+Lia%2C+Shubhashis+Roy+Dipta%2C+Abdullah+Khan+Zehady%2C+Naymul+Islam%2C+Madhusodan+Chakraborty%2C+and+Abdullah+Al+Wasif%2E+2025%2E+Read+between+the+lines%3A+A+benchmark+for+uncovering+political+bias+in+ba) 5. Usman Naseem, Robert Geislinger, Juan Ren, Sarah Kohail, Rudy Garrido Veliz, P Sam Sahil, Yiran Zhang, Marco Antonio Stranisci, Idris Abdulmumin, "Ozge Alacam, Cengiz Acar"urk, Aisha Jabr, Saba Anwar, Abinew Ali Ayele, Elena Tutubalina, Aung Kyaw Htet, Xintong Wang, Surendrabikram Thapa, Tanmoy Chak… — [SemEval-2026 Task 9: Detecting Multilingual, Multicultural and Multievent Online Polarization](https://md.research.elamur.ai/free/en/papers/naseem-2026-semeval-2026-task-9-detecting/) 6. Usman Naseem, Robert Geislinger, Juan Ren, Sarah Kohail, Rudy Garrido Veliz, P Sam Sahil, Yiran Zhang, Marco Antonio Stranisci, Idris Abdulmumin, Özge Alacam, Cengiz Acartürk, Aisha Jabr, Saba Anwar, Abinew Ali Ayele, Simona Frenda, Alessandra Teresa Cignarella, Elena Tutubalina, Oleg Rogov, Aung Ky… — [POLAR: A Benchmark for Multilingual, Multicultural, and Multi-Event Online Polarization](https://md.research.elamur.ai/free/en/papers/naseem-2026-polar-benchmark-multilingual-multicultural-multi/) 7. Habibur Rahman, Md Munam Kazi, Md Samiul Islam, Jannatul Maoya, Md Monjurul Arif, Mahir Tagwar, Mahamudul Hasan, Md Sawkat Ali, Taskeed Jabid, and Maheen Islam. 2023. An analysis of bangla tweets on social media platform for polarity detection using machine learning algorithms. In 2023 4th Internati… [find](https://scholar.google.com/scholar?q=Habibur+Rahman%2C+Md+Munam+Kazi%2C+Md+Samiul+Islam%2C+Jannatul+Maoya%2C+Md+Monjurul+Arif%2C+Mahir+Tagwar%2C+Mahamudul+Hasan%2C+Md+Sawkat+Ali%2C+Taskeed+Jabid%2C+and+Maheen+Islam%2E+2023%2E+An+analysis+of+bangla+tweets+on+s) 8. Tahmina Rahman. 2019. Party system institutionalization and pernicious polarization in bangladesh. The ANNALS of the American Academy of Political and Social Science, 681(1):173–192. [find](https://scholar.google.com/scholar?q=Tahmina+Rahman%2E+2019%2E+Party+system+institutionalization+and+pernicious+polarization+in+bangladesh%2E+The+ANNALS+of+the+American+Academy+of+Political+and+Social+Science%2C+681%281%29%3A173%E2%80%93192%2E) 9. Surendrabikram Thapa, Ashwarya Maratha, Khan Md Hasib, Mehwish Nasim, and Usman Naseem. 2023. Assessing political inclination of bangla language models. In Proceedings of the first workshop on Bangla language processing (BLP-2023), pages 62– 71. [find](https://scholar.google.com/scholar?q=Surendrabikram+Thapa%2C+Ashwarya+Maratha%2C+Khan+Md+Hasib%2C+Mehwish+Nasim%2C+and+Usman+Naseem%2E+2023%2E+Assessing+political+inclination+of+bangla+language+models%2E+In+Proceedings+of+the+first+workshop+on+Bangla+) *Updated 31 October 2026*