Research On Entity Sentiment Analysis Method Oriented Towards Financial Field

With the rapid development of intelligent information technology,more and more users invest and make decisions on financial products through the network platform,and the sentiment analysis of financial product reviews is particularly important.On the one hand,the sentiment analysis of product reviews can provide reliable reference opinions for subsequent users,which is convenient for users to consume reasonably;On the other hand,it is conducive to merchants improving financial products and providing better services.For the sentiment analysis of financial product reviews,there are still relatively few research on sentiment analysis at the aspect level,and there are many problems.(1)Lack of publicly labeled entity-level sentiment analysis datasets in the financial field Chinese;(2)The expression of proper nouns of financial entities is diverse,and some word segmentation tools are difficult to accurately segment;(3)The emotional expression in the financial field is implicit,for example,"high return platform,thunderstorm",etc.For this reason,this thesis aims at the data characteristics in the financial field,builds a financial entity sentiment analysis corpus with a certain scale,and conducts in-depth research on the emotional polarity discrimination task of a given financial entity,as well as the joint extraction task of financial entity and emotion.The two tasks were carried out on the basis of a corpus,and the specific research work is as follows:(1)Construction of financial entity sentiment analysis datasetAt present,in the field of Chinese finance,there is only the evaluation of negative financial information and subject judgment questions under the 2019 CCF big data and computational intelligence competition,which only identifies entities with negative emotional polarity in financial texts,and does not further subdivide and judge entities with positive and neutral emotional polarity.Therefore,on the basis of the evaluation of negative financial information and subject judgment questions under the 2019 CCF big data and computational intelligence competition,this thesis marks the corresponding emotional polarity of financial entities,and the scale of the constructed financial entity sentiment analysis dataset reaches 19148,which provides strong data support for the progress of subsequent research work.(2)A method for judging the emotional polarity of a given financial entityAiming at the problem that there are multiple entities in the same text in the financial field,this thesis proposes a financial entity affective polarity recognition method based on interactive attention double-graph convolutional network(ASyn Sem GCN).This method uses the pre-trained model Ro BERTa-wwm-ext to combine entities to initially represent sentences,and then establishes the interactive information representation between entities and sentences through multi-head attention.On this basis,the syntactic and semantic graph convolutional network(Syn GCN)and semantic map convolutional network(Sem GCN)are used to represent the deep syntactic and semantic sentences,and finally,the deep representation of the entity,the word level embedding representation of the entity and the sentence embedding representation are spliced,and the emotional polarity of the entity is discriminated by the fully connected layer.Experiments were conducted on a self built financial entity sentiment analysis dataset,and the experimental results showed that the method proposed in this thesis is effective in identifying emotional polarity of financial entities,with Acc and Ma-F1 values reaching 0.9091 and 0.8818,respectively.(3)The joint extraction method of financial entities and emotionsAiming at the problems of lack of given entities,long lengths of financial entities and semantic relationships between financial entities in real scenarios in the financial field,this thesis proposes a joint extraction method of financial entities and emotional polarity based on graph convolutional network.This method uses the pre-trained model BERT-wwm-ext to initially represent the sentence,and then obtains the sentence score matrix through the self-attention mechanism,and then uses the graph convolutional network to obtain the label representation of the entity,on this basis,the multi-head attention mechanism is used to obtain the interactive information representation between the entity and the sentence,and finally,the emotional polarity of the financial entity is judged through the Transformer decoding layer,and the joint extraction of the < financial entity and the emotional polarity >binary is completed.Experiments were conducted on a self built financial entity sentiment analysis dataset,and compared with the GRACE experiment,the P value and F1 value of financial entity and sentiment polarity joint extraction increased by 3.66 and 1.42,respectively.This indicates that graph convolutional networks are beneficial for capturing semantic relationships between financial entities to more accurately determine the boundaries of financial entities,thereby improving the performance of<financial entity,sentiment polarity>joint extraction.(4)Financial entity sentiment analysis systemAccording to the emotion polarity discrimination method of given financial entities and the joint extraction method of financial entities and emotions proposed in this thesis,the emotional analysis system of financial entities is designed and implemented.The system can discriminate the emotional polarity of a given financial entity,or it can directly perform< financial entity,and the emotional polarity > binary joint extraction.The system has the characteristics of beautiful interface and clear functions.