基于改进的TextRank的自动摘要提取方法

经典的TextRank算法在文档的自动摘要提取时往往只考虑了句子节点间的相似性,而忽略了文档的篇章结构及句子的上下文信息。针对这些问题,结合中文文本的结构特点,提出一种改进后的iTextRank算法,通过将标题、段落、特殊句子、句子位置和长度等信息引入到TextRank网络图的构造中,给出改进后的句子相似度计算方法及权重调整因子,并将其应用于中文文本的自动摘要提取,同时分析了算法的时间复杂度。最后,实验证明iTextRank比经典的TextRank方法具有更高的准确率和更低的召回率。

Improved TextRank-based Method for Automatic Summarization

The canonical TextRank usually only considers the similarity between sentences in the processes of automatic summarization and neglects the information of text structures and sentence contexts.To overcome these disadvantages,we proposed an improved method on the basis of TextRank,called iTextRank,by incorporating the structure information of Chinese texts.iTextRank takes some important contexts and semantic information into consideration,including titles,paragraphs,special sentences,positions and lengths of sentences,when building the network diagram of TextRank,computing the similarities of sentences and adjusting the weights of the nodes.We also applied iTextRank into the automatic summarization of Chinese texts and analyzed its time complexities.Finally,some experiments were done.The results prove that iTextRank has higher accuracy rate and lower recall rate compared with canonical TextRank.