| In the era of big data,the natural language data with an exponential growth contains rich emotional information.Fine-grained aspect-level sentiment classification can finely distinguish the emotional polarity of the sentence in different aspects,which has more important application value in scenarios such as social media and e-commerce platforms dominated by short text.Among the methods based on statistics,machine learning and deep learning,most of the existing works adopt deep neural networks and attention mechanism to learn language features,among which the methods based on Graph Convolutional Networks have good performance.However,network propagation,pooling and other operations will bring varying degrees of emotional information loss.Moreover,the complex dependencies between specific aspects and sentences as well as the rich and changeable language features of the text will also affect the classification effect and stability of the model.This thesis focuses on the above issues.Based on ASGCN model as the basic framework,an aspect-level sentiment classification model is constructed to improve the classification performance.This thesis carries out the work as follows:1.An aspect-level sentiment classification model based on context preserved capability is constructed,which is named CPGCN-MATT.Aiming at the problem of emotional information loss in the forward propagation of multilayer neural network,this thesis uses a context gating unit in the graph convolutional layer to reintegrate the useful information in the output of the previous layer.Meanwhile,in order to strengthen the key information of sentence features and weaken the influence of noise and redundant information,a multi-grained attention computing method is designed.This method makes use of both the word order features of sentences and the global grammatical features of aspect words,and calculates the semantic correlation at the phrase-level and word-level respectively,so as to make the attention assigned to each word more reasonable.Compared with the basic model,the classification performance is improved by about2% on average.2.On the basis of CPGCN-MATT,the model named CP2GCN-A-MATT is constructed by integrating the semantic information of aspect words and the semantic dependency information between words.The word masking operation will affect the extraction of complete sentence features,and the graph convolution based on syntactic dependency tree considers less semantic dependency of words,which makes the model perform poorly on datasets with poor syntactic performance.In order to make further use of the grammatical features of complete sentences and highlight the attention to aspect words,this thesis constructs an aspect enhanced grammatical feature extraction branch using the interactive attention mechanism on the basis of CPGCNMATT,and constructs a semantic dependent feature extraction branch using self-attention weight matrix.The parallel dual-branch structure can better realize the complementarity of feature information and enhance the adaptability to different corpora.Compared with CPGCN-MATT,the classification performance is improved by about 0.5% on average.3.The established model CP2GCN-A-MATT is applied to the crawled Weibo Chinese comment data,and its classification performance in practical application is evaluated. |
