基于相关子主题消解的悖向重排序方法的研究.pdfVIP

基于相关子主题消解的悖向重排序方法研究* 华松，洪宇，张剑峰，姚建民，朱巧明 苏州大学江苏省计算机信息处理重点实验室，苏州市，215006 E-mail: hongy@ 摘 要：悖向重排序是一种面向个性化信息检索的新型排序策略，其借助相悖于查询意图的描述结构，实 现倔强非相关反馈的识别和降解。目前，悖向重排序研究中的关键问题之一是如何建立极大区别于查询意 图且最具代表性的悖向意图描述。针对这一问题，本文研究一种基于相关子主题消解的悖向重排序方法， 主要研究内容包括如下四个方面：1）基于相关反馈内容切割和聚类的子主题挖掘；2 ）基于信息量和分布 特性的相关子主题判定；3 ）悖向查询意图中相关子主题的消解；4 ）基于悖向意图的检索结果重排序。实 验采用 TDT4 语料集进行测试，结果显示该悖向排序方法显著改进了检索性能，其 MAP 值提高了近 16 个百分点，P@20 和NDCG@20 分别提高了14%和12%。 关键词：信息检索；重排序；查询意图；文本分割 Research on Opposite Re-ranking Based on Relevant Sub-topic Dispelling Song Hua, Yu Hong, Jian-feng Zhang, Jian-min Yao, Qiao-ming Zhu School of Computer Science Technology, Soochow University, Soochow City, 215006 E-mail: hongy@ Abstract: Opposite re-ranking is a novel strategy to personalized information retrieval, it utilizes the description structure opposed to query intention, to achieve the recognition and dispelling of obstinate-negative feedback. At present, an important problem of the research in the opposite re-ranking is how to establish the maximum discriminative and representative opposite query intention. Aiming at the problem, this paper proposes an opposite re-ranking method based on dispelling of the relevant subtopics, and the main research includes the following 4 aspects: 1) The mining of subtopics based on the incision and clustering of the relevant feedbacks; 2) The judgment of relevant subtopics based on the amount of information and distribution character; 3) The dispelling of relevant subtopics in opposite query intention; 4) The re-ranking of search results based on the opposite intention. The experiment makes use of TDT4 corpora to test, the results present this approach improves