Research On Key Techniques Of Aspect-based Sentiment Analysis Based On Deep Learning

With the development of social network,users have published a large number of texts with sentiment polarities for different things.These texts are of great value to social and business fields by the sentiment analysis and statistics.Early studies of sentiment analysis were coarse-grained,considering that a document or sentence contained only one sentiment polarity.Aspect-based Sentiment Analysis(ABSA),witch can carry out fine-grained sentiment analysis on the aspect which aspect terms belong to,has attracted extensive attention from the academic circles in recent years.It is an important way to improve the accuracy of aspect-level sentiment analysis to accurately extract the deep semantic features of texts and to excavate the implicit relations between words.This thesis proposes the aspect level sentiment analysis model with attention-enhancing mechanism,which is mainly explored and studied from three aspects.The main research of this thesis are as follows:1.A method of extracting aspect terms based on attention mechanism is proposed.It is difficult for the existing model to effectively extract the implicit relations between words,which leads to the problems of wrong entity extraction or incomplete extraction of aspect terms.Based on the existing model,this thesis extracts the semantic features of texts through the depth model,introduces the history-aware attention mechanism witch makes full use of the historical time step information,so as to model the implicit relationship between aspect terms and improve the extraction efficency of aspect terms.2.A method of targeted aspect-based sentiment analysis based on BERT and improved attentional mechanism is proposed.BERT is used as embedded layer,based on multiple aspect terms and multiple aspects in the text to express different sentiment polarities,and aspect terms cannot effectively find corresponding aspects and sentiment polarity,leading to irrelevant information coding issues,on existing sentiment analysis model,this thesis introduced detailed aspects vector considering the context,effectively match aspect terms and aspect,improve the performance of ACD and SP.3.The online comment analysis system for restaurants is developed.Based on the above research on opinion target extraction,aspect category detection and sentiment polarity,the model is reasonably connected in to a pipeline and applied to the online restaurants review.The fine-grained analysis of all aspects of the restaurant is realized and the effectiveness of the proposed method is verified.