Abstract: In order to deeply mine the multi-modal knowledge resources of scientific research literature and enhance the supporting role in scientific research, management and decision-making, keyword extraction, semantic retrieval methods are studied by using natural language processing technology and a tobacco scientific research knowledge discovery system based on the B/S architecture is constructed. The word frequency features of TF-IDF were optimized by incorporating location information as the initial weight of nodes in TextRank. The transfer probability between nodes in TextRank was improved by integrating word frequency features and semantic features, and the literature keyword extraction was achieved by the above improvements to TextRank. Based on the keyword extraction results, time-series analysis and co-occurrence analysis were used to realize the functions of scientific research theme analysis, scientific research entity evaluation, and scientific research decision-making analysis. The vector search scheme of ElasticSearch was used to realize the semantic search of literature. The application results show that the system can provide knowledge services throughout the stages of project application, research, summary and evaluation for researchers, managers and decision-makers, and promote the improvement of scientific research efficiency and level intobacco enterprises.