随着自然语言处理技术的发展,文本摘要技术已经被广泛应用在生活的方方面面,在司法领域,文本摘要技术能够帮助司法文本实现"降维",对迅速了解案件详情,获取案件要素有很大的帮助,促使司法向信息化、智能化发展.但是现有的摘要生成模型应用在司法文本上,生成的摘要质量不尽如人意,还存在着生成重复、冗余,与现实情况不相符等问题,特别是当行为人存在多项罪名和多项判罚时,使用常见摘要生成模型生成的摘要会出现罪罚不匹配的情况.为了解决这些问题,提出基于知识增强预训练模型的司法文本摘要生成模型LCSG-ERNIE(legal case summary generation based on enhanced language representation with informative entities),该模型在预训练语言模型中融入司法知识,并结合对比学习的思想生成摘要,提高生成摘要的质量,减少出现的罪罚不匹配情况,最终通过实验证明提出的模型取得较好效果.
Judicial Text Summarization Based on Knowledge-enhanced Pretrained Language Models
With the development of natural language processing technology,the technique of text summarization has been widely ap-plied in various aspects of life.In the judicial field,the text summarization technique can be assisted by judicial texts to achieve"di-mension reduction,"which is of great help in quickly understanding case details and obtaining case elements,promoting the develop-ment of judiciary towards informatization and intelligence.However,the existing summary generation models,when applied to judicial texts,the quality of the generated summaries is unsatisfactory,and there are problems such as repeated and redundant generation,in-consistency with the actual situation,especially when the perpetrator faces multiple charges and multiple penalties,the summaries gen-erated by common summarization models will result in mismatched accusations and penalties.To address these issues,a judicial text summarization generation model based on knowledge-enhanced pre-training models legal case summary generation based on enhanced language representation with informative entities(LCSG-ERNIE)was proposed.The proposed model integrated judicial knowledge into the pre-trained language model and combined the idea of contrastive learning to generate summaries,thereby improving the quality of the generated summaries and reducing instances of mismatched accusations and penalties.Ultimately,through experimentation,it is demonstrated that the proposed model achieves good results.
text summarizationknowledge enhancementintelligent judiciarycontrastive learning
万方数据
维普
