Internet of things (IoT) contains not only large number of services with heterogeneous description but also mobile and highly resource-constrained devices. It is key issue for IoT to find suitable services efficiently and fast. This paper proposes a service discovery approach based on probabilistic topic model for IoT. The key features of this approach include: 1) using the English Wikipedia to train a topic model with high quality and semantically enrich service text description (a form of short text) to help the topic model to extract latent topics of service more effectively; 2) employing non-parametric topic model to infer latent topics of service, which reduces the training time of the topic model; 3) making full use of the latent topics of service to automatically classify service and calculate the text similarity between service request and service, which rapidly decreases the number of services for logic signature matchmaking and accelerates similarity calculation of service text description; 4) providing a logic signature matchmaking method which supports both WSDL-based and RESTful Web service. The experimental results show that the proposed method performs much better than existing solutions in terms of precision and normalized discounted cumulative gain (NDCG) measurement value.