Research on High Performance Hybrid Model for Text Sentiment Analysis

Long short-term memory (LSTM) neural networks and attention mechanisms have been widely used in the field of sentiment analysis. However, in the existing research, sentiment analysis has always had large deviations in experimental results due to incomplete feature mining. Therefore, in this paper, a hierarchical multi-attention model is proposed, which uses LSTM and Transformer encoder structures to analyze text information step by step. Firstly, convolutional feature extraction is carried out on the preprocessed text vector, then the upper-lower text correlation feature is captured by LSTM. In addition, we use the multi-head attention mechanism for further analysis, and finally the sentiment classification is output, in which the existence of multi-attention makes the model improve the model efficiency on the basis of perfect feature extraction. The proposed LSTM-Transformer framework is evaluated on datasets such as the Sentiment Corpus: Yelp13 and IMDB. Experiments show that this method has good performance.