Incorporating Similar People for Expert Finding in
Wei Zhang, Jianqing Ma, Yiping Zhong · 2009
The task of finding authoritative people within an organization has received increased interest over the past few years. To identify an expert in a specific field, expertise evidence of candidates should be collected in the enterprise corpora to represent one's knowledge and skills. Though there have been various methods proposed for evidence collecting and expertise modeling, little work has been done to collect evidence by exploring the relationship between candidates. In this paper, we build a bridge between candidates to find out people who have similar intellectual structure within the organization and then supplement an individual by incorporating the evidence collected for his similar people. Besides, a document prior based on the PageRank weight is adopted to illustrate the significance of a page so as to improve the performance of the expert finding system. Finally, we evaluate our methods on Enterprise corpora provided by TREC. Experimental results show that the incorporation of similar people and the document prior bring gains to the final results and our proposed methods have excellent performance. I. INTRODUCTION In a large organization, it is standard to employ an expert finding system to find out professional people for problem consulting or team building. Different from document retrieval, given a topic, expert finding system would return a list of candidates who are considered as authorities in this field to the users rather than return a group of related documents. Initial solutions to expert finding are mostly database based which house the expertise and skills of employees in a database so that relevant experts can be identified (7). How- ever, it is quite effort-consuming to manually create and main- tain this database. Consequently, automated methods which can analyze the corpora to collect related information for candidates and then find out relevant experts have dominated in approaches to expert finding. In general, there are two principal automated approaches to expert modeling: candidate model and document model (3). Both of these two model involve discovering related information in the corpora to model the knowledge of can- didates. However, these methods only focus on modeling the knowledge of an individual by collecting his own evidence on the basis of name recognition, little work has been done to explore the relationship between candidates and supplement one's expertise evidence by identifying his similar people who are with similar intellectual structure. For instance, there are two candidate ca1 and ca2 with similar expertise, while ca1 is less mentioned in the corpora than ca2. Using traditional expert finding methods, more evidence will be collected for ca2 than ca1 and ca2 will get a higher score than ca1 .I f we could support ca1 with evidence collected for ca2, then the ca1 would get better evaluation so as to improve the performance of expert finding. In this paper, we propose a method based on analysis over candidates' related documents and then support a candidate with evidence of his similar people. Besides, we adopt a document prior based on its PankRank to measure the sig- nificance of a document and distinguish candidates' occur- rences in important documents from occurrences in relatively unimportant documents. Finally, we evaluate the effectiveness of our methods by using CECR collection provided by Text REtrieval Conference (TREC).