Aspect-based Sentiment Analysis Method Based On Hierarchical Attention Networks

Aspect-based Sentiment Analysis(ABSA)is a fine-grained task in sentiment analysis tasks,which can determine the sentiment polarity of a target in a given sentence.It is an important research field in natural language processing.Deep learning models with hierarchical attention structures have been proven to have the state-of-the-art performance in text classification tasks.In order to further improve the performance of deep learning models in ABSA,two deep learning models with hierarchical attention structures for ABSA tasks are proposed.The main works of this paper are as follows:Firstly,to improve the performance of ABSA,this research proposes a Hierarchical LSTM Attention Network(HLAN)as the first innovation,introducing the hierarchical attention structure into ABSA firstly.HLAN is mainly composed of a vector mapping layer,a LSTM attention layer,an interactive attention layer and a prediction layer.It integrates deep learning methods such as GloVe word vector model,bidirectional LSTM network and attention mechanism.Secondly,improved on the basic idea of the HLAN,a Hierarchical Multi-Encoders Attention Network(HMEAN)is proposed as the second innovation in this research.HMEAN consists of a vector mapping layer,a GRU Multi-Head Attention jointly encoding layer,an interactive Multi-Head Attention encoding layer and a prediction layer with pooling function.Compared with HLAN,HMEAN has a certain change.It uses the position vector mapping with sine and cosine function,a Multi-Head Attention mechanism with more parameters and a more simplified bidirectional GRU network for joint coding learning.In order to verify the performance of the two innovative models proposed in this study,experiments were conducted on the public data set SemEval 2014,which proved the effectiveness of the two models.