Research On Sentiment Classification Of Comment Aspect Term Based On Graph Neural Network

With the rapid development of mobile Internet,a large number of user comments are generated every day,which often contain rich emotional information and express the user’s attitude towards different things.By analyzing the emotional tendency in the comments,the government and enterprises can make relevant strategic decisions,and ordinary users can also get valuable reference from it.Therefore,efficient and accurate sentiment analysis technique has become a popular research direction.At present,the methods for aspect term sentiment analysis mainly employ neural network to learn the emotional characteristics of aspect terms in reviews,and combine with linguistics to improve the classification effect.Long Short-Term Memory and Gated Recurrent Unit are commonly used neural networks,but they can only learn the sequence features of sentences and cannot take into account the syntactic features of sentences.It is a common optimization method to focus on close words based on location weights including linear weight and syntactic weight,which indicate the importance of other words to the emotional judgment of the aspect term.However,many researches only consider linear weight,or consider the two weights separately.In some cases,the individual linear weight or syntactic weight has limitations,and cannot accurately reflect the importance of the corresponding words.In addition,most of previous studies model the problem based on individual aspect term,ignoring the emotional relationship between multiple aspect terms in the sentence.But if consider the influence of all aspect terms in the sentence,it is hard to deal with the sentences with variable emotional polarity.Dependency syntactic information is often used in current researches,but these researches mainly model sentences based on a single aspect term,ignoring the syntactic emotional association of aspect terms.Aiming on the above problems,the main contents of this thesis are as follows:(1)By employing Graph Long Short-Term Memory network,this thesis studies the sequence features and syntactic features of sentences,so as to extract more accurate context feature of the aspect term to improve the effect of key sentiment capturing.In addition,aiming at the limitation of individual location weight,this work first calculates the initial linear weight and syntactic weight according to the relative distance between the aspect term and its context words,and then calculates the appropriate vector through the learnable parameter matrix to represent the position weight with both linear and syntactic position information.The position weight vector is finally concatenated into the corresponding word embeddings to optimize the encoding effect.(2)This thesis sets up the connection between aspect terms according to the dependency syntactic structure of sentences.If there is a direct connection between two aspect terms in the syntactic structure,they are regarded as a pair of aspect terms with emotional connection,otherwise they are regarded as having no direct connection.Then,the context features among aspect terms are extracted by Graph Attention Network,and the emotional features among aspect terms are captured by attention mechanism for final classification.Experiments on Sem Eval2014 data sets show that the proposed method has obvious optimization effect on sentiment classification.Compared with the current representative SDGCN-G,the accuracy and Macro-F1 in the Laptop domain are improved by 1.4% and 1.92% respectively.Compared with the way of considering the emotional association between all aspect terms,the way based on dependency syntactic association has some advantages in the comments with changeable emotional polarity.