Bapyn-EnKha: A Parallel English-Khasi Corpus for Machine Translation

Bapyn-EnKha is a publicly available parallel English-Khasi corpus designed to support research in low-resource machine translation. The dataset contains 40,002 aligned sentence pairs curated from WMT 2023, Tatoeba, web-based sources, and manually translated material.

The corpus includes multi-domain content spanning religious, educational, news, and conversational text. All entries were subjected to multi-stage preprocessing, including deduplication, normalization, OCR correction, and manual validation. A subset of the data was reviewed for semantic fidelity and contextual alignment with assistance from a linguistic consultant.

The dataset is released under the CC BY-NC-SA 4.0 license and is intended to support model training, benchmarking, and reproducible research for English-Khasi and related Austroasiatic language technologies.

An accompanying technical paper describing the dataset construction and validation process is included.