面向微博搜索的时间感知的混合语言模型  

Time-Aware Mixed Language Model for Microblog Search

作　　者：卫冰洁[1,2] 王斌[1] 

机构地区：[1]中国科学院计算技术研究所前瞻研究实验室,北京100190 [2]国家计算机网络应急技术处理协调中心,北京100029

出　　处：《计算机学报》2014年第1期229-237,共9页Chinese Journal of Computers

基　　金：国家自然科学基金(61070111);中国科学院先导专项课题(XDA06030200)资助~~

摘　　要：已有研究表明,时间是影响信息检索特别是微博检索的重要因素.现有的代表性工作是将时间信息作为文档先验融入统计语言检索模型,目前主要有跟查询无关和跟查询有关两种做法.这两种做法得到的模型均基于"时间越新文档越重要"这个简单假设.然而,对实际数据集进行分析发现,大多数微博查询的大部分相关文档并没有出现在最新时刻,因此上述假设并不成立.文中从这一点出发,定义这些相关文档集中出现的高峰点为热门时刻(Hot Time),并提出新假设"越靠近热门时刻,文档越重要".基于该假设,文中提出了基于热门时刻的4个系列模型(HTLMs).在此基础上,将查询无关模型看作是文档的背景时间信息而将查询有关模型看作是文档的独立时间信息,由此引入平滑思想提出混合的时间模型(MTLM).基于TREC Microblog数据的实验结果表明,HTLM模型优于现有的工作,而混合模型项对于单一模型会有进一步的提高.Previous work has shown that time is important for information retrieval tasks,especially for Microblog search.Most existing work regarded time as the document＇s prior information under language model framework with query dependent or independent style.A simple hypothesis in these work is ＂the newer the document,the more important＂.However,by analyzing the queries from TREC Microblog Track,we found that,for many queries,most of relevant documents were not published at the newest time period.These peak points were defined as hot time in our paper.Different queries have different hot time points.It sounds ＂the closer to the hot time point the document is more important＂.Based on the above new hypothesis,this paper proposed four models based on hot time points （HTLMs）.Among these models,query independent and dependent models are regarded as background and distinctive information respectively,and then a mixed time language model is proposed using smoothing technique （MTLM）.Experimental results on TREC Microblog corpus show that HTLM models outperformed current models and the mixed model can further improve the retrieval effectiveness.

关 键 词：时间感知 微博搜索 语言模型 信息检索 社交网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]