As a critical sub-task in discourse structure analysis, implicit discourse relation recognition (iDRR) is a challenging natural language processing task. Traditional approaches focus on exploring concepts and sense in discourse, which result in poor performance. This paper first systematically explores the efficiency of shallow semantic and attitude prosody-driven sentence-level sentiment information in discourse. Next, the paper proposes a simple but effective tree structure and finally investigates the efficiency of a composite kernel. Evaluation on Penn Discourse Treebank (PDTB) 2.0 shows the importance of shallow semantic and sentiment information across the discourse, and the appropriateness of the composite kernel in iDRR. It also shows that this system significantly outperforms other ones currently in the research field.