In Chinese essay, compound sentences are the majority. Recognition of relation category is screening for semantic relation of clauses in a compound sentence, and it is the key to analyze the meaning of the whole compound sentences. In a non-saturated compound sentence, the relation words are absent. So, the non-saturated compound sentence can not be classified by the features of the relation word collocation. In this work, an unbalanced corpus of non-saturated compound sentences with two clauses is taken as the research object. This study proposes a convolutional neural network for relation classification that automatically learns features from two clauses and minimizes the dependence on pre-existing natural language processing tools and language rules. The model fuses the features of relation to improve the performance. The experimental results show that the accuracy is 97% and that the proposed model outperforms the best baseline systems with sentence level features.