EFCNN: A Restricted Convolutional Neural Network for Expert Finding

Expert finding, aiming at identifying experts for given topics (queries) from expert-related corpora, has been widely studied in different contexts, but still heavily suffers from low matching quality due to inefficient representations for experts and topics (queries). In this paper, we present an interesting model, referred to as EFCNN, based on restricted convolution to address the problem. Different from traditional models for expert finding, EFCNN offers an end-to-end solution to estimate the similarity score between experts and queries. A similarity matrix is constructed using experts' document and the query. However, such a matrix ignores word specificity, consists of detached areas, and is very sparse. In EFCNN, term weighting is naturally incorporated into the similarity matrix for word specificity and a restricted convolution is proposed to ease the sparsity. We compare EFCNN with a number of baseline models for expert finding including the traditional model and the neural model. Our EFCNN clearly achieves better performance than the comparison methods on three datasets.