In order to expand the scale of manual annotated data and thereby improve model performance, we attempt to make full use of existing heterogeneous annotations to learn model parameters. We extend coupled sequence labeling model proposed by Li et al. (2015) under the BiLSTM-based deep learning framework. The neural coupled model learn its parameters directly on two heterogeneous training data, and predicts two optimal sequences simultaneously during the test phase. A lot of experiments have been conducted on the part-of-speech (POS) tagging task and the joint word segmentation and POS (WS&POS) tagging task. The results show that neural coupled approach is superior to other methods for exploiting heterogeneous lexical data, including the multi-task learning method and the traditional discrete-feature coupled model. Neural coupled model achieves higher performance on both scenarios, i.e., annotation conversion and boost the final target-side tagging accuracy by exploiting heterogeneous data.