Sentiment analysis of Chinese microblog based on stacked bidirectional LSTM

Sentiment analysis on Chinese microblogs has received extensive attention recently. Most previous studies focus on identifying sentiment orientation by encoding as many word properties as possible while they fail to consider contextual features (e.g., the long-range dependencies of words), which are, however, essentially important in the sentiment analysis. In this paper, we propose a Chinese sentiment analysis method by incorporating a word2vec model and a stacked bidirectional long short-term memory (Stacked Bi-LSTM) model. We first employ the word2vec model to capture semantic features of words and transfer words into high-dimensional word vectors. We evaluate the performance of two typical word2vec models: Continuous bag-of-words (CBOW) and skip-gram. We then use the Stacked Bi-LSTM model to conduct the feature extraction of sequential word vectors. We next apply a binary softmax classifier to predict the sentiment orientation by using semantic and contextual features. Moreover, we also conduct extensive experiments on the real dataset collected from Weibo (i.e., one of the most popular Chinese microblogs). The experimental results show that our proposed approach achieves better performance than other machine-learning models.