Research On Deep Learning-based Emotional Data Analysis Model Of Commodity Comments

With the rapid development of Internet and e-commerce,online shopping has become a part of our daily life.When shopping online,users should not only look at the description information of the product itself,but also look at the comments of other users who have already bought the product.Through emotional analysis of product reviews,users can know the true word of mouth of products and decide whether to buy.For merchants,they can adjust their sales strategies to increase their revenue by understanding the market reputation of their products.As the text of commodity review is colloquial,machine learning method relies too much on manual feature extraction,which leads to the limitation of machine learning method in emotional analysis of commodity review.The complexity of deep neural network and the feature of automatic feature extraction make deep learning method very good for emotion analysis.However,the commonly used deep learning model is essentially text classification when conducting emotion analysis,and the emotional information in the text is not used,which will lose certain accuracy.From the above perspective,emotional information in commodity reviews is integrated into the deep learn ing model,and the emotional analysis model of commodity reviews with the structure of tri-channel-convolutional neural networks(CNN)-bi-long short-term memory(BiLSTM)-emotional-multihead-attention is proposed.Crawler technology was used to capture the comment data from Jing Dong Mall,which was divided into training set and test set after data cleaning and data balance.Collected the basic emotion dictionary published on the Internet and expanded the emotion dictionary by using the mutual information algorithm of PMI points.The traditional word2 vec algorithm is improved,adding emotional information to the loss function,so that the word vector training contains emotional information,more suitable for the emotion analysis task.The part of speech vector and the dependency vector are also trained.The CNN+BiLSTM network structure is adopted to enable the model to capture the local and global information of this paper.The position vector is added to the calculation result of the convolution layer to make it contain the sequence information,giving full play to the characteristics of BiLSTM layer.Multiple emotional attention mechanism is introduced to calculate the contribution of each word to the emotional tendency of the whole text through the emotional score of each word in the input text.Through the coding implementation and tuning and optimization,the proposed model in the Jing Dong Mall review data sets on F1 value of 95.97%,and machine learning method of support vector machine,naive bayesian model and decision tree F1 value is 86.01%,82.83% and 82.95% respectively,the depth of the common learning TextRNN text classification method and TextCNN F1 value were 91.29% and 90.08%,respectively,by comparing the visible into the depth of the emotional information learning model proposed in this paper have stronger applicability in the sentiment analysis task.