| Aspect-based Sentiment Analysis(ABSA)aims at analyzing the sentiment polarity of text-specific aspects.This thesis summarizes the ABSA previous research from the artificial stage,the statistical machine learning stage and the deep learning stage,and also summarizes the existing two problems,called "Incomplete sentence representation and noise filtering problem" and "Emotional word vector solidification problem".Relying on the above background,this thesis proposes a Graph Convolutional Network(GCN)model based on Abstract Meaning Representation(AMR)and an aspect-oriented pre-training network.The former algorithm firstly proposes an aspect-oriented generation method for the AMR graph,which can be used to directly establish the semantic association related to the aspect of the context,and then designs a network combined with Bidirectional Long-short Term Memory(BiLSTM)and GCN to learn semantic graph,which focuses on realizing relational reasoning and extracting aspect-related emotional features in order to solve the "Incomplete sentence representation and noise filtering problem".Meanwhile,this algorithm also improves the ordinary GCN by proposing a label-specific and gate mechanism-based GGCN module.Based on the Pre-trained Language Model,the latter algorithm realizes the interactive fusion of text and aspect information through the Self-Attention mechanism,which is able to generate adaptive emotional features,and on this basis,combining the advantages of the former model,it also utilizes the AMR-based GGCN module to realize relational reasoning with the input of adaptive emotional embeddings,therefore it has the ability to solve both types of the "existing problems".In order to verify the effects of the two algorithms,this thesis selects three datasets to conduct experiments,including SemEval 2014 Task4 Restaurant,SemEval 2014 Task4 Laptop and Twitter.This thesis obviously shows the effectiveness of the two models by a series of experiments,including baseline comparison experiments,ablation experiments,visualization experiments and case study. |