Research On Aspect-Based Sentiment Analysis Method Based On Deep Learning

With the rapid development of the Internet,people express their views and opinions on e-commerce platforms and social software almost every day.Through sentiment analysis technology,people’s emotional tendencies towards events and commodities can be automatically obtained from massive comment texts with emotional polarity,which is of great value to many industries.However,traditional coarse-grained sentiment analysis has been unable to meet people’s needs,and people often want to gain a more comprehensive understanding of certain things or commodities.Fine-grained aspect-level sentiment analysis aims to analyze different emotions expressed by different aspects,which can help people better understand various aspects of things or commodities.In this paper,we focus on the aspect-level sentiment analysis task,and propose two deep learning models.The main work includes the following two points:(1)A local information selection mechanism(LIS)is proposed for the interaction information and location information between context and aspect information.LIS first calculates the Semantic Relative Distance(SRD)according to the positional relationship between context words and aspect words,and divides the sentence into two parts of local information using two processing methods and SRD thresholds.Then the design rule selects the appropriate local information and global information for splicing.Finally,the multi-head self-attention mechanism is used to learn the containing sentence representation,output the hidden state and classify it.Experiments on the Sim Eval 2014 and Twitter datasets show that the LIS model achieves 87.24%,82.97% and 75.58% accuracy on the three datasets,respectively.(2)Aiming at the problem that the LIS mechanism loses the original syntax and semantics of sentences when acquiring local information,and the learning of sentence representation is too simple,a Dependency-parse Tree Distance(DTD)model is proposed.The DTD model uses a syntactic dependency tree structure to redefine Semantic Relative Distance(SRD),calculates the SRD according to the positions of context words and aspect words in the dependency tree structure,and intercepts local information accordingly.The model uses LSTM network and multi-head(self)attention mechanism to learn sentence representation,and designs an interactive learning method,which can extract sentence information more deeply,and pass the output of each step through the pooling operation,and finally splicing as the final result.Sentence expressions are used for classification.The model was validated on the Em Eval2014 and Twitter datasets.The results show that the DTD model achieves 87.91%,84.68% and 75.81% accuracy on the three datasets,respectively.Both models use Glo Ve and BERT pre-trained word vector models respectively,and conduct comparative experiments on three Non-BERT models and three BERT-base models to verify the effectiveness of the models.In addition,different SRD thresholds are set for comparative experiments,and the results show that the two models have different SRD thresholds on the Sem Eval2014 and Twitter datasets to achieve optimal results.