Extracting keywords from a text set is an important task. Most of the previous studies extract keywords from a single text. Using the key topics in the text collection, the association relationship between the topic and the topic in the cross-text, and the association relationship between the words and the words in the cross-text has not played an important role in the previous method of extracting keywords from the text collection. In order to improve the accuracy of extracting keywords from text collections, using the semantic relationship between topics and topics in texts and highlighting the semantic relationship between words and words under the key topics, this article proposes an unsupervised method for mining keywords from short text collections. In this method, a two level semantic association model is used to link the semantic relations between topics and the semantic relations between words, and extract the key words based on the combined action. First, the text is represented with LDA; the authors used word2vec to calculate the semantic association between topic and topic, and build a semantic relation graph between topics, that is the upper level graph, and use a graph ranking algorithm to calculate each topic score. In the lower layer, the semantic association between words and words is calculated by using the topic scores and the relationship between topics in the upper network allow a graph to be constructed. Using a graph sorting algorithm sorts the words in short text sets to determine the keywords. The experimental results show that the method is better for extracting keywords from the text set, especially in short articles. In the text, the important topics, the relationship between topics and the correlation between words can improve the accuracy of extracting keywords from the text set.