关键词: 判决文书, 抽取式文摘, 预训练语言模型, 门机制, 单模型融合

Abstract:

In view of the large number of information points and the high degree of structure of legal judgment documents， the traditional extractive summarization method is prone to produce redundant sentences and cannot cover all the key information. A BERT （Bidirectional Encoder Representations from Transformers） based Improved Gate Dilated Convolutional Neural Network （BIGDCNN） model was proposed to deal with the above problems. Firstly， the original data was converted to sequence annotation data， then long text representation from word granularity to sentence granularity was obtained by pre-trained language model BERT. Finally， the Dilated Convolutional Neural Network （DCNN） combined with the improved gate mechanism and a single model fusion method were used to achieve low-redundancy extraction of key information in the original document while enhancing anti-interference and reducing the risk of gradient disappearance. In the automatic summarization experiment of legal judgment documents， the ROUGE-1， ROUGE-2， and ROUGE-L scores of this model were 62.85%， 46.56%， and 59.25%， which were 15.10， 15.75， and 12.97 percentage points higher than those of the mainstream model BERT+Transformer. The BIGDCNN model solves the problem of the traditional extractive summarization method， and can be efficiently used in the automatic summarization scene of legal judgment documents.

Key words: judgment documents, extractive summarization, pre-trained language model, gate mechanism, single model fusion