Sentiment Analysis Based On Part-of-Speech And Aspect

With the development of social networks,the network began to occupy an important position in people's daily communication.A large amount of emotional information such as movie reviews,product reviews,etc.,has been created.By collecting,organizing and analyzing this emotional information included the reviews,we can better understand users' behavior,which is of great help to the business operation.Most of the previous sentiment analysis work was done relying on a large amount of manual material resources.Using deep learning methods automatically process large amounts of text and analyze sentiment tendencies is a hot topic at present.After obtaining sentiment polarity on different aspects of an entity using aspect-level sentiment analysis,we can further synthesize the sentiment on all aspects of an entity and build an overall sentiment on the entity.Sentence-level and document-level sentiment analysis incorporating aspect words is also of great importance for practical applications.In this thesis,the state of research on aspect-level text sentiment analysis is outlined and summarized.For the aspect extraction problem,a fused lexical aspect extraction model and text sentiment analysis with fused aspect words is proposed in this thesis.The main work of this thesis is as follows.(1)The aspectual word extraction model with fused lexical features is proposed on the basis of the traditional BERT model for classification.The model utilizes pre-trained models,such as BERT,to obtain better semantic information of the preceding and following texts.The lexical annotation is also performed using Universal POS tags tool and the POS is vectorized using Word2 Vec.Subsequently,the lexical vectors are fed into the self-attentive mechanism layer,and the associations between the lexical features are obtained through the self-attentive mechanism.Finally,the lexical vectors are fused with word vectors and the aspectual word extraction results are obtained through the softmax layer.(2)A sentence-level sentiment analysis with fused aspectual information is proposed.The text input sequence with aspect words added is generated using BERT-SEP in the input layer,and then the weight of each aspect word is obtained through the attention layer to get the influence of each aspect word on the overall sentiment of the sentence.Finally,the sentiment classification is obtained by a five-layer CNN.The experimental results show that the proposed model in this thesis is more effective in solving the traditional sentence-level text sentiment analysis problem.Compared with the existing traditional methods,there is an improvement in F1 value.