Researchers have released PatTransBench, a 12,000-item benchmark designed for patent machine translation (PatMT) evaluation. The dataset covers Chinese paired with English, Japanese, and Korean, and includes 6,000 short-text items drawn from sentences and claims, with additional items making up the rest of the collection.
The benchmark is stratified by International Patent Classification (IPC), which helps ensure that test material spans different technical domains rather than clustering in a few areas. It also introduces a four-dimensional evaluation protocol, moving beyond simple translation accuracy to a broader assessment framework.
Because patent language is highly technical and legally precise, generic translation benchmarks often miss the specific challenges of patent text. PatTransBench is intended to give researchers a more targeted tool for measuring and improving systems that translate patent documents across these language pairs.