Long-Text Sentiment Analysis Based on Semantic Graph

The neural network of bilateral attention, especially Bidirectional Encoder Representations from Transformers (BERT), has achieved good results since its appearance, and has obtained high scores in various tasks of Natural Language Processing (NLP). However, Transformer completely adopts Attention mechanism and discards CNN and RNN, which also brings some problems. When analyzing long-text corpus, Transformer is usually limited by text length and cannot achieve good results. Due to the shortcoming, this paper proposed a long-text analyzing method called Semantic Graph Bidirectional Encoder Representations from Transformers (SG-BERT). SG-BERT obtains word features firstly, and then obtains sentence features, in corpus. After that, a semantic graph containing whole effective emotional will be generated, which can also be regarded as emotional labeling of long-text. Finally, semantic graph will be put into BERT to analysis sentiment. We experiment SG-BERT on Bag of Words Meets Bags of Popcorn (BoWMBoP) and online shopping 10 cats for sentiment analysis. In BoWMBoP and online shopping 10 cats 0~1000 length corpus, its ACC reached 96.3 and 95.9, respectively.