Deep transfer learning for social media cross-domain sentiment classification

Abstract

Social media sentiment classification has important theoretical research value and broad application prospects. Deep neural networks have been applied into social media sentiment mining tasks successfully with excellent representation learning and high efficiency classification abilities. However, it is very difficult to collect and label large scale training data for deep learning. In this case, deep transfer learning (DTL) can transfer abundant source domain knowledge to target domain using deep neural networks. In this paper, we propose a two-stage bidirectional long short-term memory (Bi-LSTM) and parameters transfer framework for short texts cross-domain sentiment classification tasks. Firstly, Bi-LSTM networks are pre-trained on a large amount of fine-labeled source domain training data. We fine-tune the pre-trained Bi-LSTM networks and transfer the parameters using target domain training data and continuing back propagation. The fine-tuning strategy is to transfer bottom-layer (general features) and retrain, top-layer (specific features) to the target domain. Extensive experiments on four Chinese social media data sets show that our method outperforms other baseline algorithms for cross-domain sentiment classification tasks.