APDDApproachforExpertFinding一个专家发现PDD方法

单击此处编辑母版标题样式,单击此处编辑母版文本样式,第二级,第三级,第四级,第五级,*,A PDD Approach for Expert Finding,Fu Yupeng,Tsinghua Univ,Backgroud,Finding a expert in an organization using corporate information,Frequently asked question,Not well addressed in research,Hot research topic in recent years,Characteristics,Build up a relationship between topics and experts via documents,Deal with a mixture of types and formats information in corporate,Integrate different kinds of expertise data,Models for matching variants of person names and name disambiguation,Expertise information recognition and extraction,QA techniques,TREC2005 Expert Finding Task,Aim:Given a query about a region,return a expert list on that domain,Resource:,Full W3C corpus,1092 candidate list,each one with a unique personid,10 training queries and 50 test queries,Resource distribution,Results,Expertise search model,How deal with topic and expert via documents？,Query-time generated model,Aggregate model,Collection,Relative Docs,Person Profiles,Topics,Extract,Experts,Experts,Topics,Extract,Construction of PDD,Person Description Document(PDD),Context information,Distance weighted functional information,Group information,Experiments,Component of PDD,Performance with different features used solely as PDD,Experiments,Effect of word pair based ranking model,Experiments,Effect of word pair based ranking model,Feature,MAP,R-prec,p10,With wp,Without wp,With wp,Without wp,With wp,Without wp,Title,0.2620,0.2479,0.3060,0.2996,0.4280,0.4140,Bold,0.0911,0.0906,0.1343,0.1297,0.1915,0.1872,Heading12,0.2476,0.2389,0.2890,0.2835,0.4220,0.4060,Anchor text,0.1290,0.1261,0.1736,0.1748,0.2688,0.2667,Comparison of wordpair-based ranking model effect on features of PDD,Experiments,Model,MAP,p10,PDD-based search model,0.2890,0.4500,Best result in the contrastive model,0.1986,0.2450,Improvement,+45.5%,+83.7,Best result using query-time generated model in TREC2005,0.2668,0.3700,Comparative best results between PDD-based search model and our contrastive model,THU vs MSRA,Context Window,Features,Title,heading12,bold,anchor text,Title,all heading info,People clustering,Structure-based extraction VS Context Vector,Future Work,Employ other resources,especially Emails,Other applications,Software search,MP3 search,Example,Thanks！,