Aspect Level Sentiment Classification with Semantic Distance Attention Networks

Abstract

Aspect-Based Sentiment Analysis (ABSA) is a kind of sentiment analysis, which extracts opinions expressed for all aspects of the entity. Compared with document-level and sentence-level sentiment analysis tasks, aspect-level sentiment analysis tasks are more granular and more challenging. Aspect-based sentiment analysis is a fine-grained sentiment analysis task, its purpose is to analyze and predict the sentiment polarity corresponding to each aspect commented in the sentence. We proposed a semantic distance attention with BERT model (SDA-BERT), which uses BERT to obtain high-quality word vectors and semantic coding and calculate the SDA value and extract the aspect semantic features. Experiments have proved that SDA-BERT has achieved excellent results on all three public data sets, and achieving accuracy rates of 86.98%, 85.45% and 86.74%, respectively. Compared with various models of machine learning and attention mechanism, the experiment has achieved better results.