Chinese Weibo sentiment analysis based on character embedding with dual-channel convolutional neural network

As one of the most compelling NLP(Natural Language Processing) tasks, sentiment analysis becomes more and more popular. In this paper, we have proposed a new method of sentiment analysis by using pre-trained character embedding with a dual-channel convolutional neural network (char-DCCNN) to comprehend the sentiment of Sina Weibo's Chinese short comments. First of all, we divide Chinese corpus into single Chinese characters which are then trained as character vectors. Characters that appear less frequently are randomly initialized. Then, the vector matrix representing the text is input into a two-channel convolutional neural network. The vector of one channel remains static (as a kind of global feature) and another is fine-tuned (as a kind of local feature) according to the input data. Finally we record the train performance, validation performance and the final average validation performance respectively to reflect the sentiment classification results. The dataset used in this paper is NLPCC2012 micro-blog sentiment analysis datasets and the reviews of sports, film, social and other fields crawled from Sina Weibo. The 10-fold cross-validation will be employed. Three experiments are done to show the good performance of our method through three aspects - embedding, activation function, channel. In all, the char-DCCNN that we have put forward improves the sentiment classification results of Weibo Chinese short comments and possesses practical significance.