MsrZhenTranslationParity is a Chinese-English translation dataset released by Microsoft in 2018, containing human evaluation results of machine translation and translation outputs, designed to allow external verification of its claim of achieving human-level translation and to facilitate future research.

The dataset contains 6 additional English translations of the WMT17 Chinese-English language pair, including Reference-HT based on human translation from scratch, Reference-PE based on human post-editing, and translation outputs from research systems Combo-4, Combo-5, Combo-6, and the online machine translation service Online-A-1710. The dataset training set contains 2,001 samples with a total size of approximately 1.79 MB, and all fields are different translation versions of the same Chinese source sentence.

Dataset Composition

The dataset mainly contains the following fields and structure:

• Reference-HT: Human translation from scratch.
• Reference-PE: Human post-edited translation.
• Combo-4, Combo-5, Combo-6: Three translation versions from research systems.
• Online-A-1710: Translation version from an anonymous online machine translation service.
• train: Training set, containing 2,001 samples, with a size of 1,797,033 bytes.

Citation

Citation information is available at this link Achieving Human Parity on Automatic Chinese to English News Translation