Mining the latent conversations which are implied in the big amount of text messages stored on one’smobile phone, is a challenging problem. They can hardly be organized by threads, due to lack of necessary metadatasuch as “subject” and “reply-to”. This paper proposes an innovative conversation recognition model based ontemporal clustering algorithms and topic detection methods. The study first clusters the text messages into candidateconversations based on their temporal attributes, and then does further analysis using a semantic model based onlatent Dirichlet allocation (LDA). In the end, the text messages are organized as conversations based on theirintegrated correlation of temporal relevancy and topic relevancy. This approach is evaluated with a real dataset,which contain 122 359 text messages collected from 50 university students during 6 months.