A short text sentiment classification method based on feature expansion and bidirectional neural network

Short text sentiment classification has a strong practical value. It also has many applications, one of which is opinion analysis. However, the traditional methods cannot effectively manage and analyze short texts because they generally contain fewer words, and their corresponding features are sparse. In this paper, we investigate the short text sentiment classification problem and propose a novel model by integrating feature expansion and the bidirectional neural network. Firstly, we use the new word discovery algorithm and the forward maximal matching to find out the new words of the short text. Secondly, we remove the deactivated words of the short text and use pkuseg for word separation to get the feature word set of the short text. Next, we use the conditional random field and the approximate lexicon (Synonyms) to expand the feature word set of a short text. Further, we use the Bi-LSTM (Bi-directional Long Short-Term Memory) to extract the features from the expanded short text. Finally, a SoftMax classifier is employed to obtain the sentiment classification results of the short text. Experiments show that this method outperforms many other methods in terms of accuracy, recall, and F1 value.