This paper first investigates two main kinds of features of Web service description language (WSDL) documents: the structure features and the reference features. Next, a novel multi-vector model for Web services is introduced, which is distinguished from the general text representation model by the explicit features of Web services. The structure features are represented by multiple vector spaces and the term weighting in the sub-vector is determined by the reference features. A method to compute the similarity between two Web services is proposed and a Web service discovery prototype system based on this new model is implemented. Finally, a Web service discovery test collection is constructed, which has 1576 WSDL documents together with incomplete relevance judgments. The experimental results on this collection show that Web service discovery based on the proposed model is more effective than based on simple vector space model of text with the confidence of 95%.