Self-Attention Neural Network for Sentiment Analysis of Multiple Aspects in Sentences

Abstract

Sentiment analysis is a task to analyze whether opinions, feelings and attitudes in sentences are positive or negative. In the aspect-based sentiment analysis which is one of methods of sentiment analysis, the aspect information which consists of an entity and an attribute included in the sentence is extracted, and the polarity is estimated under the context. In this research, we propose a neural network model based on a self-attention mechanism that identifies multiple aspect categories and identifies target phrases for each aspect category and their polarities of positive or negative under text encoding by the pre-trained language model BERT. Then, performance of the model is evaluated using the chABSA dataset prepared in the economic field document.