Research On Aspect-level Sentiment Classification Based On Deep Learning

Aspect-level sentiment classification aims to identify the sentiment polarity expressed by a sentence about a given aspect.Currently,one of the current mainstream methods for solving this task is an aspect-level sentiment classification model based on deep learning.In this thesis,after an in-depth analysis of the existing aspect-level sentiment classification models based on deep learning,these methods have the following shortcomings:First,in the aspect-level sentiment classification model,the attention model based on the neural network has better performance,but the model processes the data sequentially during the encoding process,cannot run the data in parallel,and the calculation efficiency is low.Second,the interaction between context and aspect words in existing deep learning models is not fully realized,which greatly limits the effectiveness of the model.Third,at present,most mainstream models do not fully analyze the contribution of words.The commonly used Word2 vec and Glove word embedding methods cannot express the semantics of the same word in different contexts,and the resulting word vectors cannot accurately express the semantics of the context.Based on in-depth analysis and research on the above problems,this thesis proposes an aspect-level sentiment classification model(SCBMA)based on multi-head attention(MHA).The main works of this thesis are as follows:(1)Make the most of your attention.Adopt efficient MHA and convolution operation to replace the traditional neural network to obtain the hidden state,and then through the average pooling and MHA to further realize the interaction between the context and the aspect words,the aspect-level sentiment classification model(SCBMA)based on multi-head attention is proposed.(2)Improve the quality of word vectors.The pre-trained BERT model captures complex relationships such as character-level,word-level,sentence-level,and inter-sentence to obtain the rich features of each word,providing a solid foundation for downstream tasks.The SCBMA model designed for specific tasks is applied to the BERT model(SCBMA-BERT).The BERT model can better release the performance of BERT by further fine-tuning specific tasks.(3)Experiments were conducted on 3 data sets widely used in the aspect-level sentiment classification field,and compared with several advanced aspect-level sentiment classification methods.The results show that the model experiment results proposed in this paper are superior to other comparison methods The model proposed in this paper has improved the accuracy and macro-F1 value of the experiments on data sets in different fields.