Intrinsic Knowledge Evaluation on Chinese Language Models
This work provides a new benchmark for researchers and developers to intrinsically evaluate the knowledge encoding capabilities of Chinese Language Models, addressing a gap in current downstream performance-focused evaluations.
This paper proposes four tasks (syntactic, semantic, commonsense, and factual knowledge) with 39,308 questions to evaluate the intrinsic knowledge encoded in pre-trained Chinese Language Models. The developed probes and knowledge data serve as a reliable benchmark for this evaluation.
Recent NLP tasks have benefited a lot from pre-trained language models (LM) since they are able to encode knowledge of various aspects. However, current LM evaluations focus on downstream performance, hence lack to comprehensively inspect in which aspect and to what extent have they encoded knowledge. This paper addresses both queries by proposing four tasks on syntactic, semantic, commonsense, and factual knowledge, aggregating to a total of $39,308$ questions covering both linguistic and world knowledge in Chinese. Throughout experiments, our probes and knowledge data prove to be a reliable benchmark for evaluating pre-trained Chinese LMs. Our work is publicly available at https://github.com/ZhiruoWang/ChnEval.