Aggregation Models For People Finding In Enterprise Corpora

Along with the information explosion, enterprise search which enables good knowledge management and information organization to improve the leadership capacity of the managers and the efficiency of the staff is receiving more and more attention.Expert search is one of the most valuable domains in Enterprise Retrieval area. Finding authoritative people of a given field automatically within a large organization is quite helpful in various aspects, such as problem consulting and team building.The search accuracy of the existing expert finding system is relatively low. In this paper, a novel aggregation model is proposed to solve the problem of finding authoritative people. The main contributions of this dissertation are summarized as follows:Firstly, we introduce two of the most popular expert search models:Document Model and Candidate Model and have a further discussion of their advantages and disadvantages.Secondly, we propose an Aggregation Model to solve the problem of enterprise expert search. Compared to existing Candidate Model and Document Model, in this model, various kinds of related information in the enterprise repository is assembled to model the knowledge and skills of a candidate, for instance, the profile which gives a general description of the candidate, documents related with the candidate, people with similar intellectual structure and so on. Then the candidate is modeled as a multinomial probability distribution over these collected evidence of expertise and candidates are ranked according to the probability of the topic generated by their models. When estimating this probability, we adopt language model which has been widely used in information retrieval.Thirdly, in the frame of Aggregation Model, we try to improve the existing Document Model and Candidate Model by introducing two novel methods:a weighted candidate-document association for Document Model and a resume-building method based on sliding-window and IDF filtering for Candidate Model. Then, we carry out experiments to evaluate the effect of these two methods.Finally, Similar Expert (those candidates with similar knowledge and skills) is introduced into the expert search system for the first time. We try to improve the performance of the people finding system by making use of the relationship between candidates. Accordingly, we also do several evaluation experiments to find out the effect of the introduction of Similar Expert.Experimental results on TREC benchmark enterprise corpora demonstrate that our model outperforms current state-of-the-art approaches by a large margin.