Research On Aspect-Based Sentiment Analysis Algorithm Based On Graph Neural Network

In recent years,with the rapid development of mobile internet,many apps have entered the public’s daily lives,and the content of user comments on the internet has seen an exponential growth.Comments made by users on e-commerce and social media platforms contain a large amount of textual information,which reflects the emotional tendencies expressed towards a particular product or event from different perspectives and aspects.For businesses,analyzing product reviews can help with product updates and service improvements.For the government,analyzing the public opinion on events on the internet can help with the development of relevant policies.While discourse-level and sentence-level sentiment analysis can analyze the overall emotional tendencies of a piece of text,they cannot obtain the fine-grained emotional tendencies of different entities in the text.In real-life situations,this kind of sentiment often contains more rich information and is more helpful for downstream tasks.Aspect-based sentiment analysis,as a fine-grained task of sentiment analysis,aims to obtain the sentiment polarities of various aspects in the text and has high application value.Its subtasks mainly include aspect-level sentiment extraction and aspect-level sentiment classification.In this paper,we mainly focus on aspect-based sentiment analysis and have done the following work:(1)For the aspect-based sentiment classification task,recent works have mainly used graph neural networks to encode the connection relationships and label dependencies between words on dependency parsing trees.However,dependency parsing trees are often not accurate enough because they are guided by external parsers.To address this issue,this paper proposes a dynamic multi-channel fusion mechanism based on GAT and BERT.The meanshism can adaptively adjust the fusion weights of semantic and syntactic-related information channels according to the characteristics of sentence standardization in different datasets.Experimental results show that the two-layer multi-channel dynamic fusion mechanism fully considers the complementarity between syntax and semantics,mitigating the damage caused by erroneous dependency parsing tree information to the model as much as possible.As a result,the model outperforms most strong baseline models.(2)For the aspect-based sentiment triplet extraction task,this paper designs a dynamic multi-channel fusion mechanism based on GCN and grid sequence labeling.Firstly,we exploit the information of the dependency syntax tree and the distance between words to construct the edge matrix information of various graph neural networks.Then,in order to mitigate the damage caused by erroneous dependency parsing tree information to the model,we used a dynamic fusion mechanism to integrate the edge matrix information.Finally,we utilizes the grid sequence labeling method to decode the triplet information.In addition,we deploy Deberta as our sentence encoder to enhance the model’s performance.During the training process,we use FGM(Fast Gradient Method)adversarial training to improve the model’s generalization ability.Experimental results show that our model outperforms many baseline models(GTS-BERT、BABTABSA、EMC-GCN).