基于自注意力和多通道融合的政民互动文本情感分析研究OA

A robust optimized BERT-driven bi-directional-network attention fusion fourier(RB-BTAF)model is proposed to en-hance the effect of sentiment analysis in government-citizen interactive texts,aiming to address the issues of feature extraction and sentiment depth mining in the current models.The character and word features of the government-citizen interactive texts are obtained through the dynamic encoding of robustly optimized BERT pretraining approach(RoBERTa).The self-attention mechanism is integrated into bi-directional long short-term memory(BiLSTM)network and text convolutional neural network(TextCNN)models to extract the global and local features of the government-citizen interactive texts.The integration of the two features still adopts the self-attention mechanism.On this basis,the feature vector again uses the self-attention mechanism to adjust the weight distribution,and uses the Fourier transform to convert the feature vector into frequency-domain informa-tion,so as to improve the accuracy of sentiment classification.Experiments on the government-citizen interactive text dataset show that the RB-BTAF model has more advantages than other models,and its precision,recall and F1 value reach more than 80％.