ParsProof: A Fine-grained, Error-Typed Benchmark for Persian Proofreading Tools

We present a dedicated benchmark, ParsProof, for systematic evaluation of Persian proofreading systems, featuring fine-grained annotations across 52 error types tailored to the specific linguistic properties of Persian language. Unlike most of the existing datasets, our benchmark enables detailed error-level analysis. This is the public subset of the dataset, comprising 50 sentences out of a total of 1,029.

Citation

If you use ParsProof in your research or applications, please cite the following paper:

@INPROCEEDINGS{11620914,
  author={Saaberi, Zahra and Shamsfard, Mehrnoush},
  booktitle={2026 International Interdisciplinary Conference on Artificial Intelligence: Engineering, Health, Finance and Humanities (IICAI)}, 
  title={Benchmarking Persian Proofreading Tools: A Fine-grained, Error-Typed Evaluation}, 
  year={2026},
  volume={},
  number={},
  pages={1-6},
  keywords={Tools;Modeling;Generative Pre-trained transformer;Labeling;Printing;Writing;Conferences;Proofreading tools;fundamental NLP tools;Persian;benchmarking},
  doi={10.1109/IICAI70155.2026.11620914}}