Although the automatic model updating process has been widely used in industrial recommendation systems, there are several challenges for utilizing multi-source data to improve recommendation performance, including model and engineering level. In this paper, we introduce a novel \textbf{M}ulti\textbf{-V}iew Approach with \textbf{H}ybrid \textbf{A}ttentive \textbf{N}etworks (MV-HAN) for contents retrieval in the matching stage of recommender systems. The proposed model enables high-order feature interaction from various input features while effectively transferring knowledge between different types. Moreover, the MV-HAN employs deep neural networks with a well-placed parameters sharing strategy, improving the retrieval performance in sparse types. The MV-HAN inherits the efficiency advantages in the online service from the two-tower model, by mapping all representations, including users and contents of different types, into the same space. This enables fast retrieval of similar contents with an approximate nearest neighbor algorithm. We conduct offline experiments on several industrial datasets, showing that the proposed MV-HAN significantly outperforms baselines on the contents retrieval task. Moreover, the MV-HAN is deployed in a real-world matching system. Results of Online A/B tests demonstrate that the proposed method can significantly improve the quality of recommendations.