ComplexityMT: Benchmarking the Interaction Between Text Complexity and Machine Translation

In three linesComplexityMT is a benchmark assessing how text complexity (CEFR levels) interacts with machine translation. Across 6 languages (Arabic, Dutch, English, French, Hindi, Russian), authors test 3 open-weight models, 1 closed model, and 1 commercial MT system. Findings: higher CEFR levels make translation harder; MT shifts target text CEFR levels compared to source for most languages.

Summary generated by Claude — human-verified