Because of ignoring the semantic relations between words, traditional query expansion methods cannot achieve the desired goals to expand right keywords in the nonstandard short term. Linked Data technology exploits the graph structure of RDF (Resource Description Framework) to form Linked Open Data Cloud, and provides more semantic information. In order to take into account the semantic relationships, a new query expansion method based on semantic property feature graph was proposed by combining semantic Web and graph. It used DBpedia resources as nodes to build a RDF attribute graph in which the relevance of a node was given by its relations. First, 15 kinds of semantic property weights for expressing semantic similarities between resources were obtained by supervised learning. Then, the query keywords were mapped to DBpedia resources based on the labelling properties in the whole graph of DBpedia. According to semantic features, the neighbor nodes were found out by breadth-first search and used as expansion candidate words. Eventually, the word sets with the highest relevance score values were selected as the query expansion terms. The experimental results show that compared with LOD Keyword Expansion, the proposed method based on semantic graph achieves recall of 0.89 and provides an increase of 4% in Mean Reciprocal Rank (MRR), which offers a better matching result to users.