In this paper, a deep learning approach based on sketch filling is proposed to address the issues of insufficient information and semantic gaps in natural language queries.