【作者】 宋巍； 张宇； 谢毓彬； 高汉东； 刘挺； 李生；

【Author】 SONG Wei,ZHANG Yu,Xie Yubin,Gao Handong,LIU Ting,LI Sheng Information Retrieval Lab,Harbin Institute of Technology,Harbin,150001

【机构】 哈尔滨工业大学信息检索研究室；

【摘要】 查询分类是理解互联网用户查询意图的重要手段。由于查询很短,先对查询进行扩展丰富其特征,再进行分类被证明是解决特征稀疏的重要途径,其中利用搜索引擎返回结果是最有效的方法之一。已有工作主要利用返回结果的文本,没有充分利用URL信息。本文利用互联网上人工组织的分类网页目录,设计URL分类器,通过统计结果集合中URL的类别分布预测查询类别。人工标注从搜索引擎日志中采样的约2500个查询作为测试数据。实验表明,基于URL类别的查询分类器可以获得与基于文本扩展的分类器相当的准确率,却较大地提高了查询分类的效率。将两类分类器组合得到的分类器可以获得比单个分类器更好的分类性能。特别的是,基于URL类别的分类器与原始查询相结合时,F1值可以接近基于文本扩展的分类器,而效率与基于URL类别的分类器相当,在分类性能与处理效率上进行折衷。更多还原

【Abstract】 Web query classification is a crucial way to understand user intents.Due to the shortness and ambiguity of web queries,many works classify queries based on enriched query presentations that lead to better performance.Search results are widely used for this purpose.However,most previous methods focus on making use of texts in search results but neglect the effect of URLs.This paper presents an approach util zing returned URLs for query classification.We make use of manually collected web directory on the web to train URL classifiers.The category distributions of URLs in search results are integrated to predict a query’s category. We sampled more than 2500 queries from query log and manually labeled the category of each query.The experiments on this data shows that URL based method can achieve comparative precision as classifiers based on text-expanded queries,but improve the efficiency significantly.Further,we combine both classifiers and the combined classifier outperforms each single classifier.Specially,when combining URL based classifier and query only based classifier,the F1 measure is close to classifier based on text expanded queries and the efficiency is similar to URL based classifier which trades off between effective and efficiency.更多还原