Research On Aspect Based Sentiment Analysis In Product Reviews

With the continuous expansion of the scale of online shopping,the sentiment analysis of internet product review information can not only guide consumers to make better purchasing decisions,but also help merchants understand the advantages and disadvantages of products,so as to adjust the marketing strategy of the store,both buyers and sellers have important practical significance.This thesis mainly studies product review text from three aspects,aspect extraction,Chinese of word vector model research and aspect sentiment classification.The specific innovative achievements are as follows.(1)Two feature extraction schemes and a complex aspect phrase extraction algorithm based on hierarchical relationships are proposed.Chinese product reviews lack a suitable domain dictionary and there are complex aspect phrases of multiple words.In this thesis,HowNet dictionaries are used as semantic similarities and correlations in computational aspects as semantic features of aspect extraction.Then,feature ontology and feature sentiment word set with hierarchical relationships are constructed with different kinds of product data and product review data,and they are combined with syntactic dependencies as syntactic dependency features for aspect extraction.In addition,this thesis collects and organizes more product review information and supplements experimental data for complex aspect phrase extraction tasks.In the aspect extraction experiment,the average accuracy,recall rate and F value of the four kinds of product reviews increased by 3%,3.7%and 3.5%respectively.In the experiment of phrase extraction in complex aspect,the accuracy of accurate evaluation and coverage evaluation also achieved better performance.(2)An improved Radical Enhanced Chinese Word Embedding(RECWE)model combined with attention mechanism is proposed.Aiming at the problems that the RECWE model does not reflect the different contribution of contextual words,and the different contributions of Chinese characters and their radicals and components in each word,this thesis introduces different types of attention mechanisms to the two prediction models of the RECWE model.The experimental results show that the performance of the two evaluation files increased by 2.89%and 1.04%respectively on the similarity task;the average performance of the three topics increased by 2.07%on the analog task.(3)An improved Aspect-specific Graph Convolutional Network(ASGCN)model is proposed.Aiming at the problems that the Bi-directional Long Short-Term Memory(BiLSTM)learns sentence with long text of product reviews and discards semantic information related to aspect,and different types of product reviews have different domain characteristics,this thesis proposes an improved ASGCN model The topic model based on DI LDA is used for text analysis to obtain the word distribution in different fields and supplement the semantic information on word vectors.The context vector is combined with the attention mechanism to weight the aspect words.The experimental results show that compared with the benchmark model,the accuracy of sentiment classification is improved on all four commodity datasets.