Translations Diversification for Expert Finding: A Novel Clustering-based Approach

摘要

Expert finding is the task of retrieving and ranking knowledgeable people in the subject of user's query. It is a well-studied problem that has attracted the attention of many researchers. The most important challenge in expert finding is to determine the similarity between query words and documents authored by candidate experts. One of the most important challenges in Information Retrieval (IR) community is the issue of vocabulary gap between queries and documents. In this study, a translation model based on words clustering in two query and co-occurrence spaces is proposed to overcome this problem. First, the words that are semantically close, are clustered in a query space and then each cluster in this space are clustered again in a co-occurrence space. Representatives of each cluster in the co-occurrence space are considered as a diverse subset of the parent cluster. By this method, the query translations are expected to be diversified in the query space. Next, a probabilistic model, that is based on the belonging degree of word to cluster and similarity of cluster to query in the query space, is used to consider the problem of vocabulary gap. Finally, the corresponding translations to each query are used in conjunction with a combination model for expert finding. Experiments on Stack Overflow dataset show the effectiveness of the proposed method for expert finding.