An enormous challenge for project management is to identify similar research projects accurately and efficiently among numerous proposals. To address this challenge, this paper proposes an algorithm to calculate the similarity between research projects using an improved generating method for fused word order sentence vectors based on USIF (unsupervised random walk sentence embeddings). The experimental results show that the proposed algorithm is about 15.8% more accurate than the existing approaches. The authors also propose a pre-checking algorithm by introducing a complex research cooperation graph to enhance query efficiency. The results show the pre-checking method reduces the query time cost by 96% on average.