Fine-grained Sentiment Analysis Of Commodity Evaluation Based On Deep Learning Network

Today,with the development of e-commerce platforms,more and more consumers are shopping online,so a large amount of commodity evaluation information has been accumulated.Consumers,sellers,and manufacturers are used to obtain decision aids by browsing user reviews of products.However,the current research on the evaluation text mainly focuses on the paragraph and the sentence level.In the evaluation text,there are often multiple emotional words in one sentence,so how to accurately extract the evaluation object and emotional words become a problem.To provide more fine-grained decision-making information,this paper studies the identification of commodity evaluation elements and the analysis of emotional tendency.The main work includes:1.The identification model of commodity evaluation elements based on the improved BiLSTM was proposed.Firstly,Bert was embedded in the input layer of BiLSTM to vectorize the evaluation text.Secondly,word features,grammatical features,and words after segmentation were selected,and Bert word embedding was used as the model input variables.Then the attention mechanism was overlaid on the network layer of BiLSTM to solve the semantic dilution problem of commodity evaluation text.Finally,the CRF constraint mechanism was integrated with the output layer of BiLSTM to enhance the interpretation effect of model recognition results.The entity,attribute,and emotion words of commodity evaluation elements were identified and extracted.The validity and feasibility of this method were analyzed theoretically and verified by experiments.2.A multi-channel parallel fine-particle sentiment analysis model based on CNN and BiGRU was proposed.Firstly,BiGRU was selected to replace the one-way GRU and the traditional RNN,so that the model could consider the semantic information of the text before and after the text and solve the problem of gradient disappearance.Secondly,CNN with different sizes of convolution kernels was used to extract different local feature information of text in different channels.After the extraction of the features by two models respectively,the different feature vectors were spliced together and input into the classifier,which could not only make up the defects of each model but also enhance the feature extraction ability.Fine-particle sentiment tendency analysis was carried out on the evaluation elements extracted from the commodity evaluation.The validity and feasibility of this method were analyzed theoretically and verified by experiments.