Enhancing Retrieval and Ranking Performance for Media Search Engine by Deep Learning

Deep learning has emerged as a powerful technique to uncover deep structured, nonlinear features for various information processing tasks, such as image processing, speech recognition, and information retrieval. In this paper, we introduce a framework of utilizing a Deep Structured Semantic Model (DSSM) to build similarity features to enhance search engine relevance. To illustrate the effectiveness of the proposed deep learned similarity features, we applied our method to an Xbox Media Search Engine. Specifically, we leverage a large-scale Bing web search log to train a generic DSSM. We then tune the DSSM parameters -- making the model specific to media search -- by using a training dataset from the Xbox Media Search Engine. Finally, we conduct a series of experiments in building Xbox search engine with and without the proposed DSSM similarity features. Our experiment results show that adding DSSM-based similarity features significantly improves the retrieval and ranking performance.