Aspect-based Sentiment Analysis Based On Dual-level Dependency Parsing And Implicit Sentiment Learning

With the rapid development of network applications,more and more people are using platforms such as social networks and e-commerce sites to share their perceptions of products or organizations.These perceptions often contain subjective emotional attitudes that cannot be objectively verified.However,automating the analysis of the subjective sentiment contained in these perceptions is critical for applications such as decision aids.Therefore,sentiment analysis that enables this goal has become one of the rapidly growing research areas in recent years.Traditional sentiment analysis methods mainly detect the overall sentiment polarity of documents or sentences,but the assumption of uniform emotional expression in text is not valid in reality.Therefore,Aspect-Based Sentiment Analysis(ABSA)has become a focus issue in the sentiment analysis field.Early ABSA tasks mainly focused on identifying the sentiment polarity of a given aspect,namely Aspect-level Sentiment Classification(ASC).With the development of research on jointly predicting multiple sentiment elements,and the Aspect-based Sentiment Quadruple Prediction task(ASQP)that extracts all aspectcategory-opinion-sentiment quadruples in a sentence has become a more comprehensive expression of ABSA.Existing ASC task solutions usually use graph neural networks to model syntactic dependency trees to learn feature representations of aspects,however,the variability of dependency relationships between different syntactic positions and aspects in the dependency trees makes the models vulnerable to the interference of irrelevant opinion words on aspects and may fall into semantic confusion.Not only that,the current ASQP task methods usually use label semantic information to generate target sentiment elements,which cannot yet identify the implicit sentiment attitudes in sentences lacking explicit aspect terms or opinion words.To address the above issues,the following research was conducted in this thesis:(1)To address the problems in ASC tasks,this thesis proposes a dual-level dependency parsing and local enhance based relational graph attention network.First,the definition of dependency grid and its dependency measure are given,which can portray the difference degree of dependency between aspects and words inside and outside the grid.Then,an enhanced dependency parsing method based on the dependency grid is used to refine the aspect-related dependencies by dual-level dependency parsing to isolate the interference of neighboring irrelevant opinion words to the aspect words.Finally,a Gaussian functionbased contextual focus method is used to adaptively adjust the context feature weights according to the sentence length,which can overcome the problem that the traditional local focus mechanism is insensitive to long-distance words.The model essentially isolates the negative effects of neighboring irrelevant opinion words and distal context on aspects,thus effectively solving the semantic confusion problem.The experimental results show that the model is superior in handling ASC tasks and provides a feasible idea for solving the semantic confusion problem.(2)To address the problems in the ASQP task,this thesis proposes an implicit sentiment learning based pattern transferred generative network.First,a new personalityshared framework is designed,and a supervised comparison-based implicit sentiment knowledge extraction method is used to learn domain-general shared pattern knowledge and domain-specific personality pattern knowledge,respectively.Then implicit sentiment sharing learning method is used to help implement implicit sentiment recognition on the new dataset by shared pattern transfer.Finally,an implicit sentiment-oriented aspect sentiment quadratic generation method is used to filter irrelevant information by combining variational information bottlenecks and fine-tune the generation model using structured templates.The essence of the model is to learn generic implicit expressions of sentiment from existing datasets to solve the problem that current ASQP models cannot recognize implicit sentiment.The experimental results show that the model is superior in both overall performance and implicit sentiment recognition.