Research On Fine-grained Text Sentiment Classification Method Based On Deep Learning

With the rapid development of Internet technology,there emerge a lot of unstructured text information that is massive and full of personal emotional tendencies on the cyberspace.People are eager to carry out analysis and processing of web text information through effective methods so as to explore the hidden value in it.Text sentiment classification is divided into coarse-grained sentiment classification and fine-grained sentiment classification,which aims to carry out inference and prediction on the sentiment tendencies contained in the text.In the fine-grained sentiment classification method,there are issues such as polysemy of sentiment words,lack of necessary connection between aspect words and contextual information,and uneven distribution of attention weights.This paper will focus on these issues to carry out the analysis and research on fine-grained sentiment classification method.The main research is as follows:(1)For the issue of polysemy of sentiment words,optimize the utilization efficiency of aspect words.A fine-grained sentiment classification model(ALBERT-BiGRU)was constructed that integrates the ALBERT pretrained model and the bidirectional gated recurrent neural network.The model used ALBERT pre-training model to carry out vectorized representation of text context and aspect words,combined the BiGRU network to finally obtain the hidden states of text context and aspect words,and combined the interactive attention mechanism to match the appropriate weight information for the relevant sentiment words of aspect words in text context,and the effect of sentiment classification was effectively improved.(2)For the issue that attention weights distribution is not combined with context-related information thus leading to unbalanced attention distribution.A model(IATT-BiGCN)was constructed on the basis of the ALBERT-BiGRU model that incorporates the interactive attention mechanism and syntactic dependencies.The model improved the original graph convolutional network,established a bidirectional dependency syntactic tree,and proposed Bi-directional Graph Convolutional Network(BiGCN)that can sense the directional information of the dependency syntactic tree to mine the syntactic dependency information between words and aspectual words in the text.By mining specific aspectual word representations through the mask layer,the interaction information between textual context and aspectual words was fully obtained by combining the interactive attention mechanism,and the effect of sentiment classification was further improved.(3)The model effect was validated on the SemEval public dataset.The experimental results showed that the accuracy of ALBERT-BiGRU model reached78.85% on this dataset,which can effectively improve the accuracy of sentiment classification compared with the mainstream fine-grained sentiment classification models.The IATT-BIGCN model,which incorporates the interactive attention mechanism and syntactic dependencies,achieved 81.79% accuracy on this dataset,and effectively improved the sentiment classification accuracy.