Research On Aspect-based Sentiment Analysis Using Dependency Relations And Graph Convolutional Network

With the rapid development of various network platforms,massive amounts of data texts have been generated in the fields of social networking,food delivery,live broadcasting,short video,etc.Fine-grained mining and analysis of texts have great social and commercial value.Compared with the coarse-grained sentiment analysis,aspect-based sentiment analysis aims to analyze the sentiment polarity of an entity or an attribute of an entity in a sentence.It has more general application scenario and is attracting extensive attention from academia and industry.In recent years,methods for modeling the connections between aspect and opinion words using graph convolutional neural networks and dependencies have made great progress in the field of aspect-based sentiment analysis.However,the existing research still has problems such as insufficient relative position information between words and insufficient information interaction between sentences and aspect words.Aiming at the above problems,this thesis explores and improves on the basis of existing research.The main work includes:(1)Considering that sentiment polarity prediction results are affected by the relative positions between aspect words and opinion words,this thesis proposes a aspect-based sentiment analysis model using dependency relations and rotary position embedding(AS-DRP).Using Transformer combined with rotary position embedding to capture the relative position information and semantic information between words,and and capture the syntactic information of dependencies between words through dependencies and multi-layer GCN.Then,two kinds of information are exchanged by Biaffine layers,and the gating mechanism is used for feature fusion to obtain the aspect word embedding after feature fusion.Finally,the aspect word representation is concatenated with the classification vector obtained by BERT for sentiment polarity prediction.The effectiveness and generalization ability of the model are verified by comparative experiments and case studies on public datasets.(2)Considering that the complex relationship between multi-aspect sentences and aspect words affects sentiment classification results,this thesis proposes an aspectbased sentiment analysis model(DR-Caps Net)based on dependencies and capsule networks.The complex relationship between words and sentences is modeled by the capsule network to improve the model effect.Firstly,BERT is used to obtain the word vector representation of sentence and aspect words.The multi-layer GCN extracts the dependent syntactic features of sentences,and uses the residual network to connect with the word vector of the sentence to obtain the sentence features with grammatical information.Then,the sentence features and aspect word representations are input into the capsule network,and through the aspect-aware normalization and capsule routing guidance mechanism,the most relevant representations of aspect word sentiment in the sentence are transferred to the sentiment category capsule,then the sentiment polarity prediction of aspect words is performed.Experimental results show that the DRCaps Net model achieves competitive classification results in aspect-level sentiment analysis tasks.This thesis focuses on aspect-based sentiment analysis tasks,improves the model based on the research of dependencies and GCN,and proposes the AS-DRP and DRCaps Net models,The effectiveness of the two models is verified by comparing them with the models proposed in recent years on the Laptop,Restaurant,and Twitter datasets.