Elamur Research⌕
☰
home / papers / 2025 / Empowering Bengali Education with AI: Solving Bengali Math Word Problems through Transformer Models

Empowering Bengali Education with AI: Solving Bengali Math Word Problems through Transformer Models

Paper · 2025 · Jalisha Jashim Era, Bidyarthi Paul, Tahmid Sattar Aothoi et al.

Summary

Automatically turning math word problems into equations has been studied mostly for English and Chinese. Bengali, a low-resource language, had no dedicated dataset for this task. The authors address this gap by building "PatiGonit", a set of 10,000 elementary-level Bengali word problems. Native speakers translated a subset of MAWPS problems into Bengali and localized them, adapting numerals, currency and terminology. Each item pairs a problem text with its target equation. The authors treat solving as sequence-to-sequence generation: a model outputs the equation, and a separate equation solver computes the numeric answer. They fine-tuned and compared four transformer models: a Basic Transformer trained from scratch, mT5, BanglaT5 and mBART50, while varying learning rate, epochs and batch size. The pretrained multilingual models clearly outperformed the from-scratch baseline. mT5 performed best, with mBART50 nearly matching it. The work provides the first Bengali benchmark for this task and indicates that pretrained seq2seq models transfer well to equation generation in low-resource educational settings.

Key points

  • Dataset: PatiGonit contains 10,000 problems split 80/10/10 into train, validation and test sets.
    • 9,752 problems need a single operation: 1,761 addition, 3,217 subtraction, 1,610 multiplication and 3,164 division.
    • Only 248 problems combine several operators.
  • Best result: mT5 achieved the top scores, 97.30% accuracy alongside 94.68 BLEU, using batch size 8 and 15 epochs.
  • Model comparison: mBART50 scored 97.20%, and BanglaT5 was close behind. The Basic Transformer peaked at 77.30%.
  • Source data: The problems were taken from MAWPS, which the paper describes as containing 38,000 items. Only 10,000 were translated because of resource limits.
  • Limitations:
    • The dataset is dominated by simple arithmetic, so multi-step problems are underrepresented.
    • Translation introduced ambiguities, for example with foreign names and cultural references.

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

Also on the byline: Ahsanullah University of Science and Technology.

 Fields, topics and concepts

📚 Shelves 5The 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 worksComputational Theory and Mathematics · 252 works

🧭 Topics 1Research subjects and languages the work is about.Reasoning · 1,426 works

Cite

Cite as Jalisha Jashim Era, Bidyarthi Paul, Tahmid Sattar Aothoi, Mirazur Rahman Zim and Faisal Muhammad Shah. 2025. Empowering Bengali Education with AI: Solving Bengali Math Word Problems through Transformer Models. arXiv:2501.02599.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2501.02599 · arXiv 2501.02599

@misc{era2025empowering,
  title = {{Empowering Bengali Education with AI: Solving Bengali Math Word Problems through Transformer Models}},
  author = {Jalisha Jashim Era, Bidyarthi Paul, Tahmid Sattar Aothoi, Mirazur Rahman Zim and Faisal Muhammad Shah},
  year = {2025},
  doi = {10.48550/arxiv.2501.02599},
  eprint = {2501.02599},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2501.02599}
}

References

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

  1. Patel, Arkil, Satwik Bhattamishra, and Navin Goyal. ”Are NLP models really able to solve simple math word problems?.” arXiv preprint arXiv:2103.07191 (2021).
  2. Vaswani, A. ”Attention is all you need.” Advances in Neural Information Processing Systems (2017).
  3. Bhattacharjee, Abhik, et al. ”BanglaBERT: Language model pre-training and benchmarks for low-resource language understanding evaluation in Bangla.” arXiv preprint arXiv:2101.00204 (2021). — [arXiv:2101.00204](https://arxiv.org/abs/2101.00204) find
  4. Nayak, Anish Kumar, Rajeev Patwari, and Viswanathan Subramanian. ”MathBot–A Deep Learning based Elementary School Math Word Problem Solver.” https://www.semanticscholar. org/paper/MathBot-[5] Feigenbaum, Edward A. ”Some challenges and grand challenges for computational intelligence.” Journal of the… ↗
  5. Zhang, Biao, et al. ”Variational neural machine translation.” arXiv preprint arXiv:1605.07869 (2016).
  6. Wang, Yan, Xiaojiang Liu, and Shuming Shi. ”Deep neural solver for math word problems.” Proceedings of the 2017 conference on empirical methods in natural language processing. 2017. find
  7. Chiang, Ting-Rui, and Yun-Nung Chen. ”Semantically-aligned equation generation for solving and reasoning math word problems.” arXiv preprint arXiv:1811.00720 (2018). find
  8. Liang, Zhenwen, et al. ”Mwp-bert: Numeracy-augmented pre-training for math word problem solving.” arXiv preprint arXiv:2107.13435 (2021). find
  9. Raiyan, Syed Rifat, et al. ”Math word problem solving by generating linguistic variants of problem statements.” arXiv preprint arXiv:2306.13899 (2023). find
  10. Wang, Lei, et al. ”Translating a math word problem to an expression tree.” arXiv preprint arXiv:1811.05632 (2018). find
  11. Xie, Zhipeng, and Shichao Sun. ”A goal-driven tree-structured neural model for math word problems.” Ijcai. 2019. find
  12. Chatterjee, Oishik, et al. ”WARM: A Weakly (+ Semi) Supervised Model for Solving Math word Problems.” arXiv preprint arXiv:2104.06722 (2021). — [arXiv:2104.06722](https://arxiv.org/abs/2104.06722) find
  13. Qin, Jinghui, et al. ”Semantically-aligned universal tree-structured solver for math word problems.” arXiv preprint arXiv:2010.06823 (2020). find
  14. Niyarepola, Kashyapa, et al. ”Math word problem generation with multilingual language models.” Proceedings of the 15th International Conference on Natural Language Generation. 2022. find
  15. Xue, L. ”mt5: A massively multilingual pre-trained text-to-text transformer.” arXiv preprint arXiv:2010.11934 (2020).
  16. Bhattacharjee, Abhik, et al. ”BanglaNLG and BanglaT5: Benchmarks and resources for evaluating low-resource natural language generation in Bangla.” arXiv preprint arXiv:2205.11081 (2022). find
  17. Tang, Yuqing, et al. ”Multilingual translation with extensible multilingual pretraining and finetuning.” arXiv preprint arXiv:2008.00401 (2020). find
  18. Luo, Liangchen, et al. ”Improve Mathematical Reasoning in Language Models by Automated Process Supervision.” arXiv preprint arXiv:2406.06592 (2024).

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/era-2025-empowering-bengali-education-ai-solving/