Aspect-Level Sentiment Analysis Method Based On Transformer

The need for automated sentiment analysis algorithms continues to expand as the number of comments and other sentiment-influencing texts on the Web continues to grow.Aspect-based sentiment analysis(ABSA),an important fine-grained sentiment analysis problem that aims to analyze and understand people’s perspectives at the aspect level,has been attracting great interest in the last decade.To deal with ABSA in different scenarios,scholars have introduced different tasks to analyze different sentiment elements and their relationships,including aspect terms,aspect categories,opinion terms,and sentiment polarity.Unlike earlier ABSA work that focused on a single emotion element,many composite ABSA tasks involving multiple elements have been investigated in recent years to obtain more complete aspect-level emotion information.However,previous studies suffer from the following shortcomings.1 The importance of syntactic information is not fully exploited.This makes current aspect extraction models often fail to correctly detect the boundaries of aspectual items consisting of multiple words,which affects their accuracy and reliability.Therefore,we need to adopt more efficient and accurate algorithms to fully utilize grammatical information.2 ABSA aims to identify aspect terms,corresponding emotional polarity,and opinion terms.Most research has focused only on a subset of these tasks,which has led to a variety of complex ABSA models.The following 2 main research points are proposed around the theoretical and practical application of sentiment analysis:1)To address the problem that current research generally uses a single increase in model depth and training data volume to study human syntactic information in insufficient depth,this study explores the syntactic aspects of sentences,and the study proposes the use of syntactic relative distance to weaken the adverse effects of irrelevant words,which have weak syntactic links to aspectual terms.It also uses Gated Recurrent Unit and Self-Attentive mechanisms for syntactic learning.This is expected to increase the accuracy of the aspectual sentiment classifier.Experiments demonstrate that the dependency interaction method and syntactic learning model proposed in this paper can make good use of syntactic dependencies of sentences,thus mentioning the accuracy of aspect-level sentiment classification.2)We use a text generation-based framework to handle various ABSA tasks.It uses a sequence-to-sequence text generation approach that selectively outputs the aspect words of the target,their associated sentiments and corresponding opinion items,and normalizes the results using word vectors.The use of a text generation model avoids the redundant downstream task design for different tasks,compared to the commonly used sequence annotation approach.The model proposed in this paper achieves advanced results on several datasets.An average F1 score gain of 2.68 was achieved on the aspect extraction sentiment classification(AESC)task,while the results on the aspect sentiment triad extraction(ASTE)task were more significant,obtaining an average F1 score gain of 5.03.This demonstrates the effectiveness of the generated sequence-to-sequence based approach in capturing the interaction between aspect terms,opinion terms and sentiment polarity to be than the sequence annotationbased approach.