Research On Aspect Sentiment Analysis Based On Deep Neural Network

With the rise of social networks and e-commerce platforms,more and more people are willing to post online reviews about shopping,travel,services,etc.These online texts with personal subjective emotional attitudes are used in information discovery,recommendation system construction,etc.The aspect is extremely valuable.In the field of sentiment analysis,document-level and sentence-level sentiment analysis can only mine the overall sentiment polarity information,and cannot analyze the user's sentiment and opinions on specific entities or attributes in the text.Therefore,aspect-level sentiment analysis has emerged.Is an entity or attribute.In recent years,the development and widespread use of deep learning have also provided new solutions for aspect-level sentiment analysis tasks.This article mainly studies the aspect-level sentiment analysis of web text.News text data has the characteristics of multiple coverage and wide coverage.The input word vectors of existing methods lack the problem of contextual semantic features.Combining the advantages of multi-task learning,this paper proposes a Bi LSTM-CRF-based aspect word extraction method.On the basis of the Bi LSTM-CRF network model,a language model learning layer is added,so that the model has advantages in generating semantic feature expression of word vectors and reducing overfitting to specific tasks.Finally,it is compared with other five models such as Bi LSTM and LSTM-CRF.The experimental results show that the Bi LSTM-CRF-based aspect word extraction method proposed in this paper has good performance,and the accuracy rate P,the recall rate R and the F1 value are all significantly improved.The traditional attention mechanism used by existing methods in aspect sentiment classification lacks sufficient mining of the interactive relationship between aspect words and original text,leading to the problem of information loss.This paper proposes a neural aspect sentiment classification based on BERT-LSTM-IAN Network model.The model dynamically generates the contextual semantic representation of each word through a pre-trained language model,and at the same time considers the target word's interactive influence on the context and context on the target word,and digs deeper into the feature information between them.Compared with 8 models such as SVM and TDLSTM,the classification performance of the BERT-LSTM-IAN model is the best.Based on the above methods,this paper conducts applied research on news text sentiment analysis,and compares with existing methods to verify the effectiveness of the method proposed in this paper in practical applications.