基于事件指导的多文档生成式摘要方法

摘要

multi-document summarzafion is often based on semantic analysis of texts.This method is more reasonable than the extractive method,but has to solve some problems like semantic analysis and natural language generation.This paper presented a novel abstractive method which treated event as basic semantic unit.It used events to represent the texts' topic distribution and direct multi-sentence compression to generate natural summary sentences.Experimental results on DUC compare favourably with state-of-the-art abstractive methods show that event is an ideal carrier for information and plays a positive role in multi-sentence compression.%建立在理解篇章语义基础之上的生成式摘要,在思想上相对于抽取式摘要更加合理,但在具体实现上却面临语义理解、自然语言生成等难题.提出了一种以事件作为基本语义单元的生成式摘要方法,通过对事件聚类反映篇章的主题分布,并利用事件指导多语句压缩生成自然语句构建摘要.通过在DUC标准数据集上进行评测,最终的ROUGE得分媲美目前主流的生成式方法,从而说明事件能够很好地承载篇章的主干信息,同时有效地指导多语句压缩过程中冗余信息的去除和自然语言的生成.