Research On Aspect-Level Sentiment Classification For Consumer Reviews

The task goal of Aspect-level Sentiment Classification(ASC)is to obtain the sentiment polarity of aspect words in text,which is one of the important research directions in natural language processing.To address the problem that traditional neural networks cannot accurately classify the sentiment of text from both explicit and implicit aspects,this dissertation proposes a model TS-GCN(Truncated History Attention and Selective Transformation Network-Graph Convolutional Networks).The details of the study are as follows.(1)An aspect word extraction model TS-BiLSTM was proposed based on Truncated History Attention(THA)and Selective Transformation Network(STN).TSBiLSTM improves the accuracy of aspect extraction by solving the problem that explicit aspect words and implicit aspect words cannot be extracted at the same time.First,the text is pre-trained through the Glo Ve model to obtain word embeddings containing semantic information of the text.Then,the aspect word initial representation and the opinion word initial representation are extracted through two Bidirectional Long Short Term Memory Network(BiLSTM),respectively.Then,the current aspect prediction features are calculated by inputting the aspect word initial representations into the THA model.The opinion summaries are computed by inputting the initial representation of the opinion words into the STN model.Finally,a fully-connected network is constructed using the current aspect prediction features and opinion summaries to establish the mapping relationship between explicit aspect words and opinion words.Finally,a fully-connected network is constructed using the current aspect prediction features and opinion summaries to establish the mapping relationship between explicit aspect words and opinion words.(2)TS-GCN model is proposed to combine Graph Convolutional Networks(GCN)on the basis of TS-BiLSTM model.The TS-GCN model can grasp the global information of the text using the spatiality of the graph and accomplish both explicit and implicit classification,further improving the accuracy of ASC.First,the form "[CLS] + comment + [SEP] + explicit and implicit aspect words + [SEP]" is constructed using the extracted explicit and implicit aspect words with the consumption comment utterance.The encoding of the sentences contains aspectual word information by input to the BERT model.Then,the relationships between words in the consumption comment are obtained through BiLSTM.finally,we construct the relationships between words into a text graph structure input to the GCN model for training.Words with the same aspect category undergo aggregation through the transformation of nodes in GCN.The aspect-level sentiment classification is accomplished by computing the sentiment polarity of the aggregated words and the corresponding aspect categories.The TS-GCN model was experimented on the consumer review datasets of Semeval 2014,Semeval 2015 and Semeval 2016.The experimental results show that the TS-GCN model has improved F1 values on the ASC task compared to the existing model.Table [16] Figure [26] Reference [84]...