Abstract: This paper focuses on analyzing the text sentiment tendency based on the deep learning model and starts with improving the neural network model based on public corpora to provide fine-grained analysis of text sentiment tendency and more accurate predictions. In the existing research, the extraction and utilization of text emotional features are usually based on Valence-Arousal space (VA space), but they do not pay attention to some subjective text details with emotional tendencies, such as the punctuation marks or emotional words, which could in turn to decrease the prediction accuracies made by models. Aiming at this issue, this paper proposes a hybrid Bidirectional Long Short-term Memory (BiLSTM) and Convolution Neural Network (CNN) model with an attention mechanism. Notably, in order to make our models easier to be applied to some light-weight products, we adopt the most basic components of nature language process (NLP) models. Firstly, BiLSTM is used to extract bidirectional context dependency information, and an attention mechanism is exploited to assign different weights to words that play different roles in sentiment judgment. Further, CNN is used to extract the local features of the upper layer’s output to ensure the robustness of feature extraction. The experiment shows that the combination of the methods with this order, BiLSTM first and CNN later, can achieve results which are significantly better than the existing baseline models reported in the literature.

Key words: text sentiment intensity prediction, valence-arousal space, bidirectional long short-term memory, convolution neural network, attention mechanism