thenlpresearcher/test_data_human_validated_eng_mar

Virām: Diagnostic Benchmark for Punctuation-Robust English-to-Marathi Machine Translation

Paper | GitHub | Demo

Virām (also referred to as PEM or Punct-Eng-Mar) is the first diagnostic benchmark for assessing punctuation robustness in English-to-Marathi machine translation. It consists of 54 manually curated, punctuation-ambiguous instances designed to evaluate how variations and inconsistencies in punctuation (such as missing commas) impact translation quality.

Dataset Summary

Punctuation is a vital signaling system that resolves semantic and structural ambiguity. This benchmark focuses on instances where the absence of punctuation creates natural human ambiguity (e.g., "Let's eat Grandma" vs "Let's eat, Grandma"). The dataset helps evaluate whether MT systems can preserve the intended meaning of ambiguous English source text when translating into Marathi.

• Total Instances: 54
• Focus: 38 out of 54 instances focus on ambiguity caused by a missing comma.
• Languages: English (Source) to Marathi (Target).

Dataset Structure

The dataset contains the following features:

• sent_written: The input English sentence as written (potentially ambiguous due to missing/varied punctuation).
• sent_meant: The intended meaning or context of the sentence in English.
• gt: The ground truth human-validated Marathi translation in Devanagari script.
• punct_type: The category of punctuation ambiguity being tested.

Citation

If you use this dataset or the associated research, please cite:

@article{shejole2025assessing,
  title={Assessing and Improving Punctuation Robustness in English-Marathi Machine Translation},
  author={Shejole, Kaustubh and others},
  journal={arXiv preprint arXiv:2601.09725},
  year={2025}
}