【作者】 王秉卿； 黄萱菁；

【Author】 Bingqing Wang,Xuanjing Huang Fudan University,Shanghai,200433

【机构】 复旦大学；

【摘要】 查询反馈是通过分析反馈文档来更改原始查询项。检索词在检索公式中的权重直接影响检索结果的质量,但是缺乏可靠的依据来设置检索词权重。本文将检索看作是一个二分类问题,在线性模型的框架下,把检索词的权重设置转化为线性分类模型中的参数学习问题。对于反馈的文本集合,采用生成式模型进行训练,从而确定检索词的参数。在TREC Disk 4&5测试数据上进行了实验,采用这套方法,向量模型可以获得和语言模型相当的检索性能。在TREC 7&8的测试话题上,不做查询扩展仅设置检索词的权重的情况下,MAP分别提升了7.8%和9.6%。更多还原

【Abstract】 Relevance Feedback reforms the original query based on the returned top documents.It is crucial for the retrieval performance to assign an appropriate weight to each query keyword,which lacks of reasonable approach.The retrieval task was taken as a two-class classification and the term weight setting was transferred into the parameter learning under the framework of Linear Classification Model.The generative model was adopted to estimate the parameters by taking the top returned documents as the training data.The experiment was conducted on the TREC Disk 4 & 5 test collection and the vector space retrieval model could gain an equivalent performance with the language model by our approach.On the test queries of TREC 7 and 8,with no query expansion,the MAP was raised by 7.8%and 9.6%respectively.更多还原