bokatiq/referenceless_machine_translation_evaluation

Bengali is a low resource language in natural language processing (NLP), with dialects like Sylheti, Chittagong, and Barisal being even more underrepresented. To address this, ONUBAD introduced a parallel corpus translating these dialects into Standard Bangla and English using expert translators, providing 1,540 words, 130 clauses, and 980 sentences per dialect. We focused on the Sylheti-English pair and adapted the dataset for LLM-based machine translation (MT) evaluation. We extracted the 980 Sylheti-English sentence pairs, corrected inconsistencies, and added 520 new sentence pairs, all translated by native speakers and cross-validated for accuracy, resulting in 1,500 high-quality pairs. To simulate a real-world MT evaluation scenario, we generated translations using the NLLB-200 model, recognized for its multilingual capabilities. Two native Sylheti speakers evaluated the outputs using Direct Assessment (DA) guidelines, scoring based on semantic equivalence and fluency. Scores were averaged and z normalized to reduce inter annotator variability and outliers.

Our study that uses this dataset got accepted in CLNLP 2025. The paper and code is attached for any technical reference.

Citation

If you find our dataset or code useful in your research, please cite our paper:

@article{rahman2025llm,
  title={LLM-Based Evaluation of Low-Resource Machine Translation: A Reference-less Dialect Guided Approach with a Refined Sylheti-English Benchmark},
  author={Rahman, Md Atiqur and Islam, Sabrina and Omi, Mushfiqul Haque},
  journal={arXiv preprint arXiv:2505.12273},
  year={2025}
}