Reproduce “CLUE: A Chinese Language Understanding Evaluation Benchmark” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Liang Xu, Hai Hu, Xuanwei Zhang, Lu Li, Chenjie Cao, Yudong Li, Yechen Xu, Kai Sun and 24 more

Introducing CLUE, the first large-scale Chinese Language Understanding Evaluation benchmark, provides a standardized assessment platform for NLU models. It features nine diverse tasks, a 214 GB pre-training corpus, and a diagnostic dataset, revealing an average 11 percentage point performance gap between state-of-the-art models and human capabilities.

Key insights

  • Current state-of-the-art Chinese NLU models exhibit an average 11 percentage point performance gap compared to humans across the CLUE benchmark tasks.
  • Model performance benefits from larger size, extensive pre-training data, and whole word masking strategies, while smaller models show severe limitations on complex reasoning tasks.
  • Models struggle with deep linguistic understanding, evident in their poor performance on tasks requiring monotonicity reasoning or handling Chinese-specific aspect markers, and are hampered by tokenization methods not adapted for Chinese.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.