Abstract: 【Purpose/significance】Effective keywords are extracted from the massive buffet user comment data to build core themes and topic words, helping businesses to understand user word of mouth, so as to improve the management level of the catering industry and user accurate services.【Method/process】By integrating three different keyword extraction methods, TF-IDF, TextRank and LMKE, the optimal keywords are obtained, and then the extracted keywords are subjected to semantic clustering, topic recognition topic word mining and topic weight calculation. Finally, the effectiveness of the method is verified on the collected Meituan dataset.【Result/conclusion】The experimental results show that the fusion of the three keyword extraction methods is more effective than the single keyword algorithm. The topics after clustering of text comments are: taste, dishes, environment, service, and price. The impor⁃tance of the topics is: taste 36.2%, service 22.9%, price 15.1%, environment 13.6%, dishes 12.2%. The experimental results confirm that the method can effectively recognize and construct topics and identify topic words, calculate the key content that users pay atten⁃tion to different topics, and at the same time provide a certain theoretical and technical basis for the mining and application research of themes and topic words in the catering industry【.Innovation/limitation】A method for topic identification, topic word construction and topic weight evaluation for semi-supervised semantic clustering is proposed; the disadvantage is that this experiment only uses the food buffet in Wuhan , so it's built with a limited range of topic applicability.