| With the rapid development of Internet technology,e-commerce in its pillar industries is becoming more and more attractive.In order to improve the user experience and facilitate merchants to get feedback information,various platforms provide comment areas for users to post their comments on the purchased products and shopping experience.By reading the review information,it can promote users to make reasonable shopping behaviors,and can make the merchants improve immediately.As the scale of e-commerce continues to grow,more and more comment texts accumulate in its comment area.Reading review information artificially is time-consuming and laborintensive,which will greatly discourage users from reading reviews,and will also affect merchants' access to user feedback.So how to quickly get the key information in the review is very important.Based on the deep learning method,this article can effectively extract the subject of the review and its emotional polarity through the two steps of entity recognition of the subject and viewpoint and sentiment analysis of the viewpoint.The main work and contributions are as follows:Firstly,for the word embedding part,by analyzing the shortcomings of the traditional word vector Word2 Vec,a double channels of char and word method was proposed.By using TextCNN as the classification model for experimental comparison,compared with the traditional word vector,word vector and double channels of char and word,the classification accuracy is higher.It can better retain semantic information,reduce the number of unknown words,and improve the quality of word embedding.In the entity recognition part of the evaluation subject and viewpoint,a BERT-CNN-BiGRUCRF model is proposed.BERT is used as the language model in the word embedding layer.The word vector acquired by BERT through pre training on a large number of Chinese texts contains the information of context and the semantic information of context,which can solve the problem that Word2 Vec cannot express the polysemy of a word.A convolution layer is added before the feature extraction layer in front of the general BiLSTM deep learning model.The strong ability of the convolution network to extract key semantics makes up for the feature information that BiLSTM cannot memorize too much text due to model constraints.At the same time,GRU,a variant of LSTM,is used to replace LSTM,which improves the speed of the network.This model is compared with the traditional named entity recognition model BiLSTM-CRF,and it has been improved on a number of evaluation indicators for the subject of review and the category of opinion entity.In the sentiment analysis part of the evaluation point of view,a ResGCNN model is proposed.Based on the RCNN model,the RCNN connection method of vector features is modified,and the residual network in the image processing field is applied to increase the text vector at the current position.It reduces the semantic information of the context and reduces the problem of vanishing gradients.An additional layer of BiGRU is added to enhance the extraction of context semantics.Then replaced the original "pseudo-convolution" with TextCNN,because its multiple convolution windows can obtain the n-gram feature representation in the sentence,which can better extract the key features of the sentence.And the word double-channel word embedding method proposed in this paper is applied to improve the adaptability of the model to unregistered words.The ResGCNN model with char and word channels is proposed,which has better accuracy than other text classification models in multiple data sets.Finally,the human-computer interaction of the opinion mining of the subject of product reviews was designed,and the entity recognition method based on the BERT-CNN-BiGRU-CRF model for the evaluation subject and the viewpoint was applied,and the sentiment analysis method of the evaluation viewpoint based on the char-and-word-channels ResGCNN model.This humancomputer interaction can quickly extract the evaluation subject of the product evaluation text and analyze its emotional polarity.The evaluation subject and its corresponding emotional polarity can be displayed. |
