Weibo Short-text Sentiment Classification Algorithm on Serial Hybrid Network

Aiming at the problem that the word vector representation of traditional language model can not solve the representation of polysemy, and the existing emotion classification models have some shortcomings in the extraction and combination of local and global features of short text, and lack of attention to the important information in sentences, the paper proposes a Weibo short-text sentiment classification algorithm based on BERT-CNN-BiLSTM-ATT. The word vector representation fused with text context is obtained based on the pre training model BERT; Convolutional neural network (CNN) and bidirectional long-term and short-term memory network (BiLSTM) are used to extract local and global features in series, put them into the attention mechanism to obtain important parts, allocate attention weights of different sizes, and improve the classification accuracy. Finally, the obtained feature vectors are sent to softmax function for training and classification. The experimental results show that the algorithm has the highest evaluation index in the Weibo short-text test set, and the accuracy is as high as 93.78%. Compared with Word2Vec pretraining model and other hybrid neural networks, it shows that it can effectively fuse and represent multiple features, has better short-text sentiment classification effect, better evaluation criteria and great practical value.