Event-driven Deep Learning-based Implicit Sentiment Analysis

Sentiment analysis aims to detect the sentiment polarity of a given text,which will help e-commerce platforms provide users with personalized services,optimize enterprise product strategies,and manage social public opinion.Sentiment analysis can be divided into explicit sentiment analysis and implicit sentiment analysis according to whether the text contains obvious sentiment words.Currently,most sentiment analysis methods focus on explicit sentiment analysis.However,in daily life,people use objective statements or rhetoric to express their emotions implicitly.Implicit sentiment analysis is an important part of sentiment analysis.Discovering the implicit sentiment in the text can avoid misjudging user sentiment or formulate inappropriate product strategies,further promoting the application of sentiment analysis related fields and the rapid development of the industry..Existing implicit sentiment analysis methods can be mainly divided into two categories,metaphor-based implicit sentiment analysis and event-centered implicit sentiment analysis.However,the former method using metaphorical dictionaries over-simplifies implicit sentiment analysis;the latter,although emphasizing the importance of events,does not explicitly model event information.As a result,implicit sentiment analysis cannot fully utilize event information.Therefore,this thesis focuses on the event-driven implicit sentiment analysis,and explicitly introduces event information based on deep learning technology.The main work is as follows:1.To solve the problem of lacking event information annotation in the existing ordinary sentiment analysis datasets and implicit sentiment analysis datasets,an implicit sentiment analysis dataset EveSA is constructed based on FrameNet semantic dictionary to annotate event information.Each text contains the following three types of annotations:the event triplet<subject,predicate,object>,the event type label,and the sentiment polarity label.2.Aiming at the problem of event information-driven implicit sentiment analysis,this thesis proposes an implicit sentiment analysis method PipISA based on event representation learning.The event triplet is fused by Tensor Composition mechanism to compute the event representation.Based on this event representation,this thesis combines sentences for implicit sentiment analysis.At the same time,a multi-task learning framework is introduced for joint learning of event classification and emotion classification,and the model is further focused on the sentiment contained in events through the constraint of the distance between the text representation and the event representation in hidden space.Experimental results show that the PipISA model outperforms other comparative methods on the proposed dataset EveSA and public dataset SemEval 2017 Task 4 Subtask A.3.Aiming at the problem that PipISA relies on annotated event triplet or event triplet extracted by other methods,an implicit sentiment analysis method EEISA based on multi-task learning is proposed.The fine-tuned BERT model is used to simultaneously perform three tasks of event triple extraction,event classification and sentiment analysis.Through the parameter sharing of multi-task learning,the implicit sentiment analysis task is assisted by event information,thereby improving the performance of sentiment classification.The experimental results show that the EEISA method achieves a 1%improvement in accuracy compared with the sentence-level sentiment classification method,and achieves similar performance compared with PipISA.