Research On Aspect Sentiment Classification Based On Attention Mechanism

With the booming development of the Internet and the rapid popularization of mobile smart devices,a large number of views and comments have been generated by users on social media and e-commerce platforms.These views and comments are mainly text-based and gradually presented in the form of multimodal data,such as text and image.These contents usually state different angles of a hot event or multiple attributes of a popular product,expressing users’ distinctive positions and rich sentimental information.Aspect-level sentiment classification can capture fine-grained sentiment information from these contents and more accurately identify users’ sentiment tendencies towards specific aspects,which has important practical implications for applications such as policy development,product improvement,and psychotherapy.This paper focuses on aspect-level sentiment classification.The main research contents and contributions are as follows.(1)To address the issues of insufficient extraction of semantic information and high computational complexity of attention mechanisms in existing textual aspect-level sentiment classification,a model based on a context and graph attention network is proposed.After reconstructing dependency trees with aspects as roots for mining aspect-related dependency syntactic relations,the model employs two graph attention networks and one context attention network to extract syntactic structure information and semantic information.In addition,a syntactic attention mechanism based on syntactic relative distance with low computational complexity is proposed,which can reduce computational complexity and effectively highlight the words related to aspects in syntax.Experiments on three public sentiment datasets show that the proposed model is effective.It can make better use of semantic information and syntactic structure information to improve the accuracy of sentiment classification.In particular,the context attention network and the syntactic attention mechanism can significantly improve the performance of sentiment classification.(2)To address the issues of low utilization of multimodal information and insufficient extraction of deep information in existing multimodal aspect-level sentiment classification,a model based on a multi-selection attention mechanism is proposed.Explicitly considering the relevance of images to target aspects,the multi-selection attention mechanism makes full use of shared features and private features of image modality to enhance the sentiment expression of target aspects.On this basis,a simple and effective residual encoder-decoder is proposed to fully mine deep information and avoid vanishing gradients and network degradation.Experiments on two public sentiment datasets show that the proposed model is effective.The multi-selection attention mechanism can better utilize image information to complement the semantic information of text modality.And the residual encoder-decoder can fully extract deep sentiment information.Both of them can improve the accuracy of sentiment classification.