A parallel processing method for long-range contextual semantic information to sentiment analysis based on aspect

Abstract

Aspect-based sentiment analysis is crucial for Internet applications such as social networks and e-commerce, where the previous deep learning methods cannot process long-range semantic information in parallel. This paper proposes an aspectbased sentiment analysis method based on multiscale convolution and a double-layer attention mechanism. The technique uses pre-trained BERT to obtain the hidden semantic information of the context from the training set, then uses multiscale deep convolution and double-layer attention to process the long-distance semantic information between the target word and the context in parallel, and finally uses softmax for sentiment classification of the target word. In this paper, we use the public dataset of SemEval 2014 and the Twitter Dataset to validated the improved accuracy and F1 of the model.