Compositionality, which enables the natural language to represent complex concepts using the combination of simple parts, allows us to convey an open-ended set of messages using limited vocabulary and grammar rules. Hence researchers expect the emergent communication protocol invented by the neural agents also have similar properties. Inspired by the language evolutionary procedure and the account proposed by the linguists, we propose an effective neural iterated learning algorithm to enhance the compositionality of the emergent language. We also propose a probabilistic explanation to articulate the mechanisms of the algorithm, which can also interpret many findings in related works. Using experimental results and ablation studies, our proposed algorithm are proved to be effective in facilitating the emergence of high-compositional languages, which can performs much better than low-compositional languages in zero-shot experiments.