Fine-Grained Sentiment Analysis Combined With Boundary Information

Aspect based sentiment analysis is a fine-grained sentiment analysis task,aiming to analyze users’sentiment polarity towards a certain attribute of the product with product reviews.Specifically,it can be divided into two subtask:aspect term extraction and sentiment analysis.Aspect term extraction is the core subtask between them and has been one of the performance bottlenecks for a long time.There are two points which are extremely important among many factors that affect the extraction effect.Firstly,the noise information in the comment text is far more than the attribute description,which greatly interferes with the discovery of aspect terms and influence the recall rate;Secondly,aspect term extraction is prone to boundary errors which influences the extraction accuracy.Therefore,this thesis proposes aspect term extraction based on scope detection and boundary control mechanism to solve the corresponding two problems respectively.In addition,end-to-end aspect based sentiment analysis has gradually become a new trend in this field.In this thesis,a fine-grained sentiment analysis method combined with boundary information is proposed by combining the above two points,and Prompt is also applied to further optimize the extraction quality.Specifically,the main research content of this thesis includes three parts:(1)Aspect term extraction based on scope detection.Aiming at too much noise information in product reviews,this thesis proposes an aspect term extraction method based on scope detection to distinguish which part of the review owns attribute description,so as to filter the noise information.The scope detection module estimates the approximate position of attribute description,so as to reduce the difficulty of aspect term recall and make the model pay more attention to the extracted core information.Based on the span-based extraction model,this thesis designs a two-stage joint learning method of scope detection and aspect term extraction.The first stage is joint learning scope detection and aspect term extraction,and the second stage is extracting aspect term using known attribute description.In addition,this thesis also takes the Segment Embedding of the pre-training language model BERT as the method to perceive the scope information.(2)Aspect term extraction based on boundary control.Aiming at the boundary error problem that often occurs in the process of aspect term extraction,this thesis proposes a boundary control method for aspect term extraction based on the double affine attention network,which re-estimates the boundary of the predicted attribute words to alleviate the boundary error.The joint learning of aspect term extraction and boundary control is used to help the model learn more about aspect boundary information,and the boundary of aspect term is calibrated through bidirectional boundary control,which employs ensemble learning to improve the confidence and accuracy of aspect terms.In addition,this thesis extends the method to a multi-head structure to further improve the robustness of the extraction results.(3)Aspect based sentiment analysis combined with boundary information.Aiming at the problems of error propagation in traditional pipeline aspect based sentiment analysis methods,this thesis researches end-to-end aspect based sentiment analysis methods,and proposes a aspect based sentiment analysis method combined with boundary information which is based on scope detection and boundary control.And the core ideas of the two methods are incorporated into the pre-training task of pre-training language models with Prompt learning.Finally,a public opinion analysis system for product aspect is built based on aspect based sentiment analysis.The proposed method is put into practice in the industrial scene of Taobao reviews to turn the academic achievements into industrial outputs.In summary,this thesis proposes scope detection and boundary control to help the aspect term extraction.The recall rate of aspect term extraction is improved by filtering noise information with scope detection,the accuracy is imporved for the boundary error of extracted attributes is suppressed by boundary control.This thesis also explores the end-toend aspect based sentiment analysis task and constructs an aspect based sentiment analysis method combined with boundary information.Prompt learning is applied to incorporate related knowledge into the pre-training task of pre-training language model and furtherly improve the extraction results.Finally,a system based on aspect based sentiment analysis is built to complete the academic landing.