Research On Aspect-based Sentiment Analysis With Opinion Terms

Aspect-based sentiment analysis(ABSA)aims to determine the sentiment polarity of a review on a given aspect term(also named opinion target).For example,in the sentence "the screen is not good while the battery is OK",the aspect term are "screen" and "battery",and the corresponding sentiment polarity are negative and positive.the Opinion term,also called sentiment words,refer to words or phrases that express subjective emotions in the text,such as "not good" and "OK" in the above example.Intuitively,opinion terms are very important for determining sentiment polarity.However,most of works on ABSA pay less attention to opinion terms.Therefore,this dissertation focuses on opinion terms of ABSA.The main work of this dissertation is as follows:1.Target-oriented opinion words extraction(TOWE).Most of the previous work take the opinion term(word)extraction as an auxiliary task,rather than regarding the opinion term as an important feature in sentiment classification.In order to extraction opinion terms for a specific target,we focuses on TOWE task,and proposes an extraction model based on BERT.For opinion term extraction,in addition to sequence labeling,we also try the boundary prediction method,which is commonly used in extractive question answering.The experimental results show that the proposed model is far superior to the existing method(7%-8%)in f1 measure,and the boundary prediction method is slightly better than the sequence labeling in f1 measure,and have a faster training and decoding speed.2.aspect-level sentiment classification with opinion terms(ASCO).After obtaining opinion terms,we studies the mechanism of opinion terms in aspect-level sentiment classification(ASC).Two indicators are designed to quantify the contribution of opinion terms in ASC.Experiments on the two quantitative indicators show that as the training progresses,the model will pay more attention to opinion terms.In order to learn this attention more directly,we modify the self-attention layer of BERT model.Experiments show that with our modified BERT to indicate the opinon terms of the textm,the performance of ASC is improved and the impovement of the accuracy score is up to 1.45%.3.End-to-end ABSA with opinion terms.Previous end-to-end ABSA model only output the targets and the corresponding sentiment polarity,without opinion terms.Forthe sake of completeness of ABSA results,we focuses on the task of aspect sentiment triplet extraction(ASTE),in which a triplet is composed of a target,the corresponding opinion terms and the sentiment polarity.In the existing work,ASTE is modeled as two stages.In stage one,targets and opinion terms are extracted.In stage two,they are coupled to form candidate pairs pool and then sent to a classifier to identify the legitimate ones.This paper proposes an end-to-end model for ASTE,extracting targets and opinion terms among with their relationship.Experiments show that the proposed model is superior to the existing methods on multiple indicators in multiple datasets,and the f1 measure on triplets is increased by up to 13%.