Personalized Semantic Matching for Web Search

In recent years, a wide variety of neural networks have been introduced for Web search to calculate the semantic relevance between search queries and Web pages. Although fairly good performance has been achieved, a severe drawback which impedes the existing models delivering superior performance is their one size fits all fashion: the neural networks treat the information needs from different users exactly in the same way. In order to obtain better relevance estimation between search queries and Web pages for each individual user, one novel approach is proposed in this paper to insert a user component for user embedding into the neural network for personalization and three strategies are introduced based on the insert location. These strategies incrementally add a user component to non-personalized neural networks and result in highly personalized relevance computation. The personalized neural networks do not only provide quality calibrated result for the chosen users but also guarantee no side-effect to the others. Comprehensive experiments have been carried out on a large-scale data set of a major commercial search engine to evaluate the proposed method. Experimental results verify the validity of neural network personalization and the superiority of proposed strategies.