Service clustering has played a great role in some fundamental service computing domain, such as service discovery, service selection and service recommendation. Inferring topics from short service descriptions is an important but challenge task for service clustering. BTM (Biterm Topic Model) performs well in discovering latent relevance semantic from short text. But choosing two words indiscriminately to form a biterm may affect the accuracy of topic modeling, because the distinct words are not equally indicative in presenting service topics. Therefore, a SP-BTM is proposed in this paper which only chooses the words with specific parts-of-speech to form biterms for topic modeling. By some experiments on real world dataset, it is verified that nouns, verbs and adjectives are benefit to topic representation and the SP-BTM can improve the effectiveness of service clustering.