The ability to accurately detect opinion expression in a document is an essential and fundamental task in opinion mining. In this work, we consider opinion expression detection as a sequence labeling task. We describe deep neural network frameworks that consist of convolutional neural networks (CNNs) and bidirectional gated units (Bi-GRUs). CNNs are capable of capturing local features in a sequence, while Bi-GRUs, a type of recurrent neural network (RNN) variant, are able to extract features from sequence data. The properties of these two networks provide the framework to effectively detect opinion expression. Experimental results show that our methods significantly outperform traditional methods like conditional random field (CRF) and previous state-of-the-art deep RNN methods.