Abstract: 【Purpose/significance】The essence of keyword extraction is to find the key words that can express the core semantic infor‐
mation of the document, so using semantics instead of words for analysis is more in line with the actual needs. Based on TextRank
word graph model, this paper proposes a keyword extraction method based on semantic clustering by using semantic instead of word
analysis.【Method/process】The word embedding trained with HowNet semantic information are clustered to gather the words with simi‐lar meanings and obtain the corresponding semantic categories for each word. Then, the window co-occurrence frequency of the se‐mantic category of the word is taken as the transition probability between words to calculate the node score. At the same time, the TF-IDF value and the node score are weighted together to correct the keyword extraction results.【Result/conclusion】From the keyword extraction results, the keyword extraction method proposed in this paper has a certain improvement in the extraction effect. Compared with TextRank algorithm, the accuracy P, recall R and F values are improved by 12.66%, 13.77% and 13.16% respectively.【Innova‐tion/limitation】This paper analyzes the relevance network based on semantics, and measures the weight of words through semantic clustering. This idea can be combined with the existing methods based on word analysis to replace words from the semantic level. At the same time, this paper introduces the word vector fusing the semantic information of HowNet for keyword extraction for the first time, which can be used as a reference for other natural language processing work because the keyword extraction method in this paper relies on Chinese knowledge base HowNet, it is only suitable for Chinese text extraction, and does not have the universality of cross language application.