With the advent of the Internet era, search engines begin to be widely used. In the case of unpopular data, the search engine is unable to retrieve the required data due to the small range of the user's search term. At this time, the query extension system can effectively assist the search engine to provide the reliable services. Based on the query extension method of global document analysis, a semantic relevance model which combines the neural network model with the corpus containing semantic information was proposed to extract semantic information between words in a deeper level. This deep semantic information can provide more comprehensive and effective feature support for the query extension system, so as to analyze the extensible relationship between words. The local extensible word distribution was extracted from the semantic data such as thesaurus and language knowledge base "HowNet" sememe annotation information, and the local extensible word distribution of each word in corpus space was fitted to the global extensible word distribution by using the deep mining ability of the neural network model. In the comparison experiment with the query extension methods based on language model and thesaurus respectively, the query extension method based on semantic relevance model has a higher query extension efficiency; especially for the unpopular search data, the recall rate of semantic relevance model increases by 11.1 percentage points and 5.29 percentage points compared to those of the comparison methods respectively.