Research And Application Of Multi-tuple Fine-grained Sentiment Analysis

The internet is filled with copious textual data,harboring invaluable sentiment information with significant commercial and referential value.In contrast to traditional coarse-grained sentiment analysis,fine-grained sentiment analysis extracts more specific information,such as aspect words and opinion words.This dissertation focuses on two cutting-edge multituple fine-grained sentiment analysis tasks:Aspect Sentiment Triplet Extraction(ASTE)and Structured Sentiment Analysis(SSA).The primary research areas are as follows:First,In the ASTE task,recent relevant studies have employed machine reading comprehension architectures to obtain aspect words and corresponding opinion words and sentiments through multiple rounds of questioning,achieving impressive results.However,these methods face difficulties when a sentence contains multiple aspect words that interfere with each other,thereby hindering accurate identification of sentiment information for each aspect.To overcome these challenges,the dissertation proposes a Context-based Masking Machine Reading Comprehension(COM-MRC)framework,consisting of three parts:masked data augmentation,interactive discrimination models,and staged inference methods.By synergizing these three components,the COM-MRC achieves advanced performance,validating its effectiveness.Second,in the SSA task,recent related studies typically convert the problem into a bi-lexical dependency parsing problem.However,due to overlapping and non-contiguous entities in the task,these transformation methods are inequivalent.To surmount these challenges,the dissertation introduces a bi-lexical dependency parsing method capable of uniformly addressing overlapping and non-contiguous entities.The method comprises two parsing edge types:relation prediction and word extraction,which correspond to the resolution of overlapping and non-contiguous entity issues.Furthermore,the bi-lexical dependency parsing method is transformed into a unified 2D table-filling mechanism called the USSA mechanism.This mechanism divides the table into lower and upper triangles,corresponding to relation identification and word extraction,respectively.Finally,a model compatible with the USSA mechanism is designed,whose bi-axial attention mechanism captures row-column correlation information of relation types in the table for more accurate identification.The experimental results demonstrate the USSA’s advanced performance and effectiveness.Third,based on the research of ASTE and SSA tasks,a fine-grained sentiment analysis system has been developed.The system supports multiple features,such as registration,login,user management,and online processing,making it user-friendly and accessible.