L3Cube-MahaParaphrase Dataset

Paper: MahaParaphrase: A Marathi Paraphrase Detection Corpus and BERT-based Models
Code: https://github.com/l3cube-pune/MarathiNLP

Overview:

The L3Cube-MahaParaphrase Dataset is a Marathi paraphrase detection corpus.It is a high-quality, human-annotated corpus specifically designed for Marathi, a low-resource Indic language. It contains 8,000 sentence pairs labeled as either Paraphrase (P) or Non-paraphrase (NP). This dataset is useful for tasks like paraphrase detection, semantic similarity, and data augmentation, as well as improving NLP models for low-resource languages.

Language:

• Primary Language: Marathi (Low-resource Indic Language)

Dataset Size:

• Number of Sentence Pairs: 8,000
  • Paraphrase (P): 4000 pairs
  • Non-paraphrase (NP): 4000 pairs

Annotation:

Each sentence pair in the dataset is manually annotated by human experts. The labels include:

• Paraphrase (P): Sentences that convey the same meaning with different wording.
• Non-paraphrase (NP): Sentences that do not convey the same meaning.

Intended Use:

The dataset is ideal for training and evaluating NLP models for:

• Paraphrase Detection
• Textual Similarity
• Data Augmentation for Low-resource Languages
• Transfer Learning for Indic Languages

Model Benchmarks:

Standard transformer-based models like BERT have been evaluated on this dataset, providing a performance baseline for future research.

Citation:

If you use this dataset, please cite the original work as follows:

@article{jadhav2025mahaparaphrase,
  title={MahaParaphrase: A Marathi Paraphrase Detection Corpus and BERT-based Models},
  author={Jadhav, Suramya and Shanbhag, Abhay and Thakurdesai, Amogh and Sinare, Ridhima and Joshi, Ananya and Joshi, Raviraj},
  journal={arXiv preprint arXiv:2508.17444},
  year={2025}
}