Summary: The most critural challenge of implicit discourse relation recognition lies in how to represent the semantic information of each discourse argument. However, the semantic value of the sentence is mainly decided by its specific information focus in linguistics. Therefore, the discourse relation may mostly depend on links between information focuses. Intuitively, we cannot give equal treatment to every phrase branches during composition up the syntactic parse tree. To resolve the problem, we introduce the tree-structured long short-term memory (Tree-LSTM) network to selectively incorporate information from each child to compute the distributed semantic representation of two arguments. Consequently, it can emphasize those informative predicative branches that indicate the “focus” of a sentence. Then the neural tensor network (NTN) is used to predict the semantic correlation between these two discourse arguments across multiple dimensions. Experimental results on PDTB corpus show that our model has achieved some improvement on the task of discourse relation recognition.