A Lightweight Chinese Summarization Generation Method Based on Knowledge Distillation

In recent years, large-scale pretrained natural language processing models such as BERT, GPT3 have achieved good results in processing tasks. However, in daily applications, these large-scale language models usually exist large model size or long running time, and the model portability and application are not very convenient. To settle this problem, we come up with a lightweight summarization generation method based on knowledge distillation (LW-BERT-KD, abbreviated as LBK model) to prove that complex neural network knowledge (taking BERT model as an example) can be extracted to lightweight language processing models (taking BiLSTM model as an example), thus achieving the purpose of complex knowledge extraction and model compression. In this paper, the BERT model is regarded as the teacher model, and the BiLSTM is regarded as the student model. The knowledge distillation technology is the transfer of knowledge from teacher models to student models, so that the generated model is effective, lightweight, and easy to port. On the Chinese abstract generation dataset, the training effect of this model is significantly better than the basic Transformer baseline. The experimental results show that if the training parameters are increased by 100 times, the model achieves comparable results with complex preprocessing language models.