The service description contains limited application scenario information. It makes the component Web API recommendation of Mashup service based on function matching different from the requirement expectation. Although some researchers use the collaboration relationship of Web APIs to improve recommendation compatibility, they ignore the negative feedback effect of functional association on Mashup service creation, which limits the improvement of recommendation quality. To address this problem, a Web API recommendation method for Mashup service components by fusing latent allied words and association compatibility is proposed in this paper. Firstly, a service description feature word extraction model S-YAKE with scenario adaptation weights was constructed, and the latent allied words were extracted by combining label adaptation. The function vectors of fusing the latent allied words were generated for Mashup service requirements and Web APIs respectively. Secondly, the heterogeneous service association graph was established, and the association vector of Web API was trained by the improved GATNE model network. Finally, the attention mechanism was introduced to optimize and recombine the Mashup requirement vector and the Web API association vectors. They are fed into the fully connected layer together with the Web API feature vectors to achieve Top-K component Web API recommendation based on the output probability of the matching degree. Compared with the state-of-the-art methods, the proposed method improves the recall and precision by 7.96% and 8.34%, and reduces the ILS by 13.07% on average. The recall and precision values of cold-start Web API recommendation are 46.97% and 49.1% of those of non-cold-start Web API recommendation, respectively. It indicates that the proposed method not only improves the quality of Web API recommendation, but also has a good recommendation effect on cold-start Web API.