Xlnet Parallel Hybrid Network Sentiment Analysis Based on Sentiment Word Augmentation

Abstract

Aiming at the problem that some sentiment words in sentiment analysis tasks are not highly concerned and it is difficult to capture the long-distance dependence between sentences, a parallel mixed sentiment analysis network EXLNnet-BG-Att-CNN that integrates the characteristics of sentiment words is proposed. First of all, the basic emotion dictionary used is HowNet, Dalian University of Technology Chinese Emotion Dictionary Database, and National Taiwan University's Emotion Dictionary, combined with the additional expansion of related emotional polarity words, to obtain a more targeted emotional dictionary and design An emotional word selection segmentation algorithm (DicSentencePieceSelect) is proposed. Secondly, use XLNnet to encode the sentence and the vectors processed by the dictionary word selection algorithm respectively to obtain the deep semantic features of the text and merge them. Then, the feature vectors are input into the parallel network of BiGRU-ATT and TextCNN respectively, and the dual attention mechanism is used, which can not only take into account the advantages of the global features of the text sequence, but also further extract local features to achieve semantic enhancement. Finally, the output vectors of the various networks are fused, and the activation-pooling layer is used to avoid the occurrence of over-fitting. Compared with multiple existing models, the accuracy rate of Acc is higher, and the accuracy rate of the model reaches 96.05%.