Sentiment Classification Model of Online Reviews Based on Word Features and Bi-LSTM

With the rapid development of e-commerce, many purchase and comment records are produced. Sentiment classification of commodity reviews is of great value for automatically monitoring bad reviews and assisting merchants in analyzing consumer feedback. At present, the Bi-LSTM model is representative of Chinese text sentiment classification, which can understand the semantic information in time sequence. However, due to the lack of processing lexical information, there is a problem that word vectors cannot highlight the information of sentiment words. Therefore, this paper proposes a sentiment classification model of Chinese product reviews based on word features and Bi-LSTM. The new model firstly uses Word2vec's CBOW model to train the word vectors, secondly uses an improved information gain algorithm with the word distribution and sentiment weights to calculate the amount of information, and finally uses the Naive Bayes model to classify the network classification results twice, which solves the problem that the basic Bi-LSTM model lacks understanding of lexical information. The experimental results show that the new model achieves better results relative to the basic Bi-LSTM model and can capture the sentiment information of the comments more accurately. In the test set, the accuracy rate reached 89.03%.