相似度质心多层过滤策略的动态文摘方法  

The similarity centroid multilayer filtering dynamic summarization method

作　　者：于洋[1,2] 范文义[1] 刘美玲[1,2] 王慧强[2] 

机构地区：[1]东北林业大学林学院,黑龙江哈尔滨150040 [2]哈尔滨工程大学计算机科学与技术学院,黑龙江哈尔滨150001

出　　处：《哈尔滨工程大学学报》2014年第10期1236-1241,共6页Journal of Harbin Engineering University

基　　金：国家863计划资助项目(2012AA102001);国家自然科学基金资助项目(60736014);中央高校基本科研业务费专项资金资助项目(2572014CB26)

摘　　要：为了研究网络快速有效获取信息的方法,网络动态演化内容的识别和分析成为人们迫切需要解决的关键问题。动态多文档文摘建立在时间信息基础上,从网络数据的动态性能入手,对同一主题不同时段的文摘集合进行分析,在识别信息内容差异性的基础上,对信息的动态演化性进行建模。在提出相似度累加模型基础上,进一步提出了基于质心整体选优的动态文摘模型。分析当前文档集合与历史集合强关联性,以选择出的不同文摘句为首句生成候选文摘集合,然后根据质心多层过滤优选方法从中选出最优文摘结果。这种模型方法消除了因首句选择不当而对文摘性能造成的影响,在国际标准评测Taxt Anynasis Conference 2008的Update task任务语料上进行了测试,并且获得了较好的实验结果。To research the method for quickly obtaining effective information on the internet,identifying and analyzing dynamic evolution of the network has become a key issue that needs to be resolved urgently. Dynamic multidocument summarization is based on the time information starts from dynamic performance analysis of network data,the analyzes the abstracts collect about the same topic in different periods of time,and the models of dynamic evolution of information on the basis of identifying differences of information contents. This paper first introduced the text similarity cumulative model and then the dynamic summarization model based on centroid integer selection. The high relevance between the current collection of documents and the historical collection was analyzed and different sentences summaries were selected and used as the first sentences of candidate set of abstracts newly generated.Next,the best abstracts were selected from the results based on the centroid multilayer filtering optimization method. These models eliminate the impact on the abstract performance due to poor choice of the first sentences. Experiments on the update task corpus from the Taxt Anynasis Conference 2008（ TAC2008） were conducted and the comparison results between new models and TAC2008 evaluation showed the effectiveness of the dynamic summarization models.

关 键 词：动态文摘模型 质心整体选优 相似度累加模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]