How to automatically generate or recommend a set of Web APIs for Mashup creation according a user's natural language description of requirement is a focus of attention among business process managers and services composition designers. A topic adaptive Web API recommendation method, HDP-FM (hierarchical Dirichlet processes-factorization machine), is proposed in this paper to recommend a set of Web APIs for Mashup creation. This approach firstly makes the Web API description document as a corpus, and trains a topic distribution vector for a Web API by the HDP model. It then predicts a topic distribution vector for a Mashup via the generated model, where the topic distribution vector is used to calculate the similarity. Finally, a factorization model is utilized to score and sort Web APIs by taking the similarity between Mashups, the similarity between Web APIs, the popularity of Web APIs and the co-occurrence of Web APIs as inputs. A Mashup can be created based on these recommended Web APIs. To verify the performance of the HDP-FM method, a series of experiments are conducted on a real dataset crawled from the ProgrammableWeb platform. The results show that the HDP-FM method has a good performance over others in term of precision, recall, F-measure and NDCG@N.