Research On Precision Classification Technology Of Appeal Short Text

With the development of community grid management,the number of citizen appeal gradually shows a blowout growth trend.The classification of massive appeal texts can effectively reduce the working pressure of civil service hotline staff,improve the efficiency of civil service hotline,and realize the good interaction between government governance and citizen participation.How to effectively classify and manage massive citizen appeal has become one of the hot research topics.In this thesis,we use the methods based on BERT model and ERNIE model for text classification task in view of the crawled dataset of short appeal texts.In order to extract the deeper semantic features in text information,this thesis uses the BERT-Bi LSTM model to classify short appeal texts.This method uses BERT pre-training models to obtain feature vectors containing context semantic information,and inputs them into the Bi LSTM model to encode and fuse the sequences to obtain the final appeal category.In order to further improve the classification accuracy of short appeal texts,this thesis integrates the key information affecting the appeals in the appeals text into the input representation of the model in the form of manual annotation,and proposes the ERNIE-AA model integrating the artificial attention mechanism.This method introduces the artificial attention mechanism into the input vector of the model,and obtains the feature vector containing the key information of the category through the ERNIE pre-training model for classification.The experimental results show that the text classification model based on BERTBi LSTM extracts deeper semantic features,and the classification accuracy is 88.17 %,which is 1.18 % and 0.34 % higher than that of BERT model and ERNIE model,respectively.More importantly,the ERNIE-AA model proposed in this thesis obtains semantic features containing key information through artificial attention mechanism,and the classification accuracy can reach 91.01 %,which is 2.84 % higher than that of BERT-Bi LSTM model.The classification model of short appeal texts based on ERNIE-AA has obvious advantages in classification effect.In addition,experiments have proved that the ERNIE-AA model also has excellent effects on rhetorical classification tasks,and the ERNIE-AA model has generalization in text classification tasks in other fields.