Research On Optimization Of Aspect Representation Methods In Aspect-based Sentiment Classification

The massive network data generated by the Internet creates opportunities for the development of text sentiment classification tasks.Traditional text sentiment classification tasks use paragraphs or sentences as basic units to judge the overall sentiment tendency of paragraphs or sentences.Aspect-based sentiment classification aims to determine the sentiment tendency(negative,positive or neutral)expressed by a specific aspect in a given sentence.At present,there are two common problems in the research of aspect representation methods in the field of aspect-based sentiment classification: the learning process of the aspect representation vector is independent of the sentence encoding process,which makes the model unable to determine the aspect meaning according to the sentence semantic;while introducing noise information(such as words with no actual meaning in aspects),it is inevitable to lose other important information.Therefore,this paper uses the optimized aspect representation method as the starting point to conduct related research on aspect-based sentiment classification.The main research contents are summarized as follows.1.This paper designs a mutual enhanced mechanism,and builds a METNet(Mutual Enhanced Transformation Network)model based on this mechanism.The METNet model contains two important components,namely the aspect enhancement module and the aspect-specific transformation unit.The aspect enhancement module incorporates sentence information into the aspect representation vector in a feature fusion manner,and can obtain a context-aware aspect representation vector.The aspect-specific transformation unit connects the context-aware aspect representation vector with the representation vector of each word in the sentence for nonlinear conversion,and obtains the aspect-aware sentence representation vector.It can be seen that the two components cooperate with each other to realize the mutual enhancement of the sentence representation vector and the aspect representation vector.2.This paper constructs a aspect-based sentiment classification model based on graph convolutional network and modified aspect representation vectors,referred to as GCN-TMA(Graph Convolutional Networ—Transformation Based on Modified Aspect Representation Vectors)model.This model also contains two important components,namely the graph convolution module based on dependency tree and the aspect-aware context transformation unit.The graph convolution module based on the dependency tree first constructs the corresponding graph structure based on the dependency tree of the sentence.The nodes of the graph represent words,and the undirected edges represent whether there is a dependency relationship between the words corresponding to the two nodes.Then,the status of the current node is updated based on the neighboring nodes,and the node update is implemented through convolution operation.Finally,a masking operation is performed on the output of the convolutional layer to retain the aspect representation vector that integrates syntactic information.The aspect-aware context transformation unit first calculates the attention score of each word in the sentence with the aspect.Then,different aspect representation vectors are generated for each context word based on the attention score.Subsequently,the encoding representation vector of the context word and the corresponding aspect representation vector are connected for non-linear transformation,and the aspectaware sentence representation vector can be obtained and finally used to calculate the sentiment classification result.This paper verifies the effectiveness of the METNet model and GCN-TMA model on the Sem Eval 2014 dataset and Twitter dataset.Compared with the interactive sentiment classification model IAN,the classification accuracy of the METNet model on given datasets has increased by 4.5% on average.Compared with the CoattentionMem Net model based on alternating coattention,the classification accuracy of the METNet model on given datasets has increased by 4.1% on average.Compared with the ASGCN-DG model based on the graph convolutional network,the classification accuracy of the GCN-TMA model on given datasets is improved by 2.8% on average.Compared with the TNet-LF model that uses modified aspect representation vectors,the classification accuracy of the GCN-TMA model on given datasets has increased by 2.4% on average.