The invention provides a term extraction method and apparatus for solving the problem that the demand of the term extraction method based on statistical information for the target corpus scale is relatively large, but the relatively small corpus scale usually causes unreliable statistics of low-frequency words. The method comprises the following steps: extracting nouns and noun phrases in first corpus to serve as a candidate term set T; obtaining embedded expressions of candidate terms according to encyclopedia corpus and the candidate term set T; calculating the semantic similarity among thecandidate terms according to the embedded expressions of the candidate terms; with the candidate terms as vertexes, constructing a margin between a candidate term a and a candidate term b by using thesemantic similarity SR (a, b) among the candidate terms to obtain a term communication network G=(E, V), wherein V represents a vertex set of the term communication network G, and E represents an undirected edge set of the term communication network G; sorting the candidate terms according to the term communication network by using a confidence propagation algorithm; and screening the terms according to the sorting. By adoption of the term extraction method and apparatus, the term extraction performance is greatly improved.