Aspect-based Sentiment Analysis Based On BERT And Interactive Attention Mechanism

Aspect-based sentiment analysis is an important research task in the field of natural language processing,which can mine the sentiment viewpoint of specific aspect(e.g.,price,service)in text at a fine-grained level,and classify the sentiment polarity.The research task can provide consumers with guidance on consumption decisions,and provide theoretical support for businesses to develop marketing strategies.For the existing deep learning methods,semantic information contained by the model input vector is limited and the perception of aspect words is weak.To solve the problem,the dissertation proposed an aspect-based sentiment analysis model based on BERT and interactive attention mechanism(BERT-Bi GRU-IATT).The main research contents are as follows.(1)An aspect-based sentiment analysis model based on BERT and Bi GRU(BERTBi GRU)is proposed to explore the deep semantic features of texts,and solve the problem of shallow representation of model input vector.Firstly,the BERT pretraining model is used to encode sentences and aspect words respectively to generate word vector representation containing rich semantic information as model input.Then,the bidirectional GRU network is used to learn the contextual features of words from two directions,and the hidden layer feature vectors are obtained and spliced to form a complete vector representation.Finally,the aspect word representation and context representation are concatenated as the final representation,which is input to Softmax function for sentiment classification.Experimental results on Sem Eval 2014 Task 4dataset verifies the validity of BERT-Bi GRU model,the average accuracy and Fmeasure are 79.79% and 74.64%,respectively.(2)Furthermore,the interactive attention mechanism is introduced into BERTBi GRU model to capture the interactive information between aspect words and context,which can solve the problem that the key information in the context is not focused on.Firstly,the interactive attention mechanism is added into the neural network training process,and the context and aspect word feature vectors output from Bi GRU hidden layer are performed by average pooling.Then,in virtue of the average representation of hidden layer,different attention weights are assigned to context and aspect words respectively in an interactive computing way to highlight key information.Compared with the BERT-Bi GRU model,the accuracy and F-measure of BERT-Bi GRU-IATT model are improved by 1.34% and 1.23%,respectively.Experimental results show that BERT-Bi GRU-IATT model proposed in this dissertation can accurately judge the sentiment polarity corresponding to aspect words on the Sem Eval 2014 Task 4 dataset.The aspect-based sentiment analysis task can be widely applied to intelligent recommendation and question answering system,which has potential application value for users and businesses in the field of e-commerce.Figure [22] Table [10] Reference [82]...