Heterogeneous information networks can be used for modeling several applications in the real world. Their representation learning has received extensive attention from scholars. Most of the representation learning methods extract structural and semantic information based on meta-paths and their effectiveness in network analysis have been proved. However, these methods ignore the node internal information and different degrees of importance of meta-path instances. Besides, they can capture only the local node information. Thus, this study proposes a heterogeneous network representation learning method fusing mutual information and multiple meta-paths. First, a meta-path internal encoding method called relational rotation encoding is used, which captures the structural and semantic information of the heterogeneous information network according to adjacent nodes and meta-path context nodes. It uses an attention mechanism to model the importance of each meta-path instance. Then, an unsupervised heterogeneous network representation learning method fusing mutual information maximization and multiple meta-paths is proposed and mutual information can capture both global and local information. Finally, experiments are conducted on two real datasets. Compared with the current mainstream algorithms as well as some semi-supervised algorithms, the results show that the proposed method has better performance on node classification and clustering.