Fine-grained Sentiment Analysis Of E-commerce Review Text Based On Deep Learning

With ever-changing Internet technologies and the advent of the 5G era,people tend to browse through goods and buy various daily supplies on the Internet,where they can also express their views and feelings on the goods.These comments contain various subjective opinions of users,and thus their sentimental tendencies can be identified by analyzing the comments and effective improvements of commodities can be made.At present,problems still exist in the sentiment analysis of e-commerce review texts,such as inaccurate sentiment expression of word vectors and coarse-grained sentiment analysis due to lack of extraction of evaluated objects.The main tasks of this paper are as follows:(1)To tackle the problems that the word vectors generated by the Skip-Gram model cannot distinguish words with opposite sentimental tendencies and differences of the same word under various parts of speech,this paper proposed a generation method of WSV-POSSO(Word Sentiment Vector based on Part of Speech and SO-PMI)which can integrate Skip-Gram,parts of speech,and SO-PMI algorithm.In this method,a Senti Table was firstly constructed to represent the word vectors under different parts of speech.And then SO-PMI algorithm was used to obtain the sentiment information of words.Finally,the author applied the improved Skip-Gram model to train the sentiment word vectors of the text.After the part-of-speech information was added in the training process,the generated word vectors could distinguish different parts of speech of the same word.After the sentiment information was added,the synonyms and antonyms of words could be distinguished more clearly without affecting the semantic expression of the words.According to the experimental results for IMDB and NLPCC2014 datasets,the sentiment word vectors generated by the proposed WSV-POSSO method are more accurate in sentimental classification than those by methods of Skip-Gram,GLOVE,and SSPE.(2)The current conditional random field method for commodity comment extraction requires a large number of artificial features;does not factor into the correlation between the input and output;and fails to fuse the context in the full-text range.To tackle the above problems,this paper proposed a method to extract opinion target for sentiment analysis which can fuse the attention mechanism and BI-LSTM-CRF.First,with the sequence labeling method of BIO,this paper turned the opinion target extraction task into a sequence labeling task,and then used the bidirectional long short-term memory network to process sequence labeling.The attention mechanism was later introduced to calculate the probability distribution of attention allocation and improve the expression ability of sequences.Finally,the conditional random field was adopted to calculate the optimal labeling paths of the text sequences,thus improving the accuracy of evaluated object extraction.Experiments on the datasets of evaluated object extraction in Sem Eval-2014 showed that this method has higher extraction accuracy than methods such as CRF,LSTM,and BILSTM.(3)In order to solve the problems of low classification accuracy and weak generalization of a single classification model in sentiment analysis of comment texts,the author proposed a stacking(ensemble learning)model based on LDA(Latent Dirichlet Allocation)dataset partitioning.The model first intensified the differences among base classifiers based on LDA dataset partitioning,and then sentiment word vectors were used to embed words into the training sets to train different base classifiers.Finally,the model integrated the classifiers using logistic regression and multi-layer perceptron.In the end,the paper used sentiment word vectors and the ensemble learning method based on LDA dataset partitioning to conduct sentiment analysis on the extracted sentiment objects,thus completing fine-grained sentiment analysis.This method effectively intensifies the differences among various classifiers and improves the classification accuracy.