LightBERT: A Distilled Chinese BERT Model

Abstract

AbstractPre-trained language models (e.g. BERT) have achieve remarkable performance in most natural language understanding tasks. However, it’s difficult to apply these models to online systems for their huge amount of parameters and long inference time. Knowledge Distillation is a popular model compression technique, which could achieve considerable model structure compression with limited performance degradation. However, there are currently no knowledge distillation methods specially designed for compressing Chinese pre-trained language model and no corresponding distilled model has been publicly released. In this paper, we propose LightBERT, which is a distilled Bert model specially for Chinese Language Processing. We perform pre-training distillation under the masking language model objective with whole word masking, which is a masking strategy adapted to Chinese language characteristics. Furthermore, we adopt a multi-step distillation strategy to compress the model progressively. Experiments on CLUE benchmark show LightBERT could reduce 62.5% size of a RoBERTa model while achieving 94.5% the performance of its teacher.KeywordsModel compressionKnowledge distillationLanguage model