Research And Application Of Corporate Public Opinion Analysis Technology Based On Multiple Fusion Neural Network

The goal of corporate public opinion analysis is to provide users with emotional information in the text collection of the company,to understand the target Company's legal risk,News sentiment,Corporate details,etc.,so that users can accurately grasp the portrait of the target company.In view of this,this thesis mainly researches the corporate sentiment analysis task,news topic classification task,corporate information extraction task and court information extraction task in the corporate public opinion scene,and combines the text matching technology to realize the application process of each task.In the study of corporate sentiment analysis technology,this thesis proposes a multifused neural network model(TFRCNN),which can better understand the corporate sentiment in the text by learning the semantic information of sentences of corresponding length.A structure is designed in the model which uses feature fusion extraction and model deep fusion.The feature fusion extraction structure uses attention encoder and residual fusion network for data learning,which achieves the purpose of in-depth analysis of sentence structure through multiple feature fusion methods.The model's deep fusion structure improves the short sentence analysis effect by constructing three-layer RNN and CNN combination units,and ultimately improves the accuracy of the sentiment analysis of the enterprise.The experiments in this thesis on different sentiment analysis datasets show that the model has better results than those models such as convolutional neural network(CNN)and long-short-term memory(LSTM)combined with attention mechanism.In addition,this thesis also applies the TFRCNN model to short text matching tasks and entity information extraction tasks.Experiments show that this model has better results than the contrast models.Finally,based on the TFRCNN model,this thesis constructs a corporate public opinion analysis system and a data review labeling system.The specific work is as follows:1.In view of the shortage of news topic classification and corporate sentiment analysis data in the corporate public opinion system,this thesis proposes a data augmentation method combining with the automatic data labeling process.Based on this,a data review labeling system was developed.This system can be used to review the labeled data online and correct the data that was processed automatically by mistake.After processing with this method,a high-quality news topic classification and sentiment analysis data set was finally obtained in the corporate public opinion analysis scenario.Based on this well-processed data set,This thesis determines the optimal models for each algorithm task in the public opinion analysis scenario.2.The requirements analysis,system design,core functions and business logic of the two systems are described in detail.The corporate public opinion system mainly includes modules such as target corporate information retrieval and query,news retrieval,corporate sentiment information query,news topic category and popularity index query,news category and source data visualization,court announcement file uploading,announcement analysis results displaying,Which provides the functions of corporate public opinion analysis.The data review and labeling system mainly includes automatic data labeled,news topic classification data review,and sentiment analysis data review,so this system contains the functions of data review and verification.Finally,the two systems in this thesis combine Flask technology with the mysql database to implement various business functions.