Research On Petition Correction And Identification Algorithm Based On Reinforcement Learning

With the widespread popularity of the Internet and the vigorous development of big data technology,our government service platforms actively integrate with the Internet and constantly improve the working mechanism of absorbing public opinion and gathering public wisdom.At present,the online government petition platform has been gradually improved,the number of participants has increased,and a huge amount of petition text data has emerged.Rapid and accurate analysis of social opinion and public concerns reflected by petition has important decision-making value for promoting social governance.Manual identification of petition expectations is inefficient,so it is of strong research and practical application significance to develop algorithm and online platform for automatic petition expectation identification.This paper takes the largest online petition platform "Message board for leaders" as an example,and carries out the research on the automatic identification of petition expectations.The online petition platform mainly relies on users’ manual labeling to obtain requests,which is subjective,inefficient and highly mislabeled,making many requests that really need to be responded to in a timely manner be ignored or submerged.The research of petition expectation identification faces challenges such as insufficient contextual information,unbalanced text structure,non-standardized petition text,and noisy labeled data.To address the above challenges,this paper models the petition expectation identification problem as a multi-classification task from the perspective of improving the efficiency of government response,i.e.,by analyzing petition textual information and predicting the urgency level of response(Urgency,Suggestion,and Gratitude).In this paper,we propose a novel framework Pecid RL for petition correction and expectation identification based on deep reinforcement learning to correct noisy labels in petition datasets and build multiple petition expectation identification models to obtain better petition expectation identification accuracy.The main research work and contributions of this paper are as follows:(1)This paper extracts multi-view petition short text features,including word-level and document-level semantic features,short text structure features based on graph representation learning,sentiment analysis features,and fusion features integrating semantic features and structure features.(2)The reinforcement learning module is introduced to address the noisy label problem correcting the wrong and fuzzy labels.The experimental results show that the reinforcement learning module for label correction proposed in this paper has excellent noise correction performance with an average improvement of each experimental metric of 8.3% and the highest improvement rate reaching 14.2%.(3)By fusing multi-view text features with traditional machine learning models,classical deep learning models,and graph neural network,19 petition expectation identification models are extended and the classification performance before and after label correction is compared.The results of the experiments show that better classification performance are obtained based on the label-corrected data.Among them,the Peti-SVM-bert has the highest identification accuracy of 93.66%,which is decided as the final petition expectation identification model.(4)An online web-server is developed for Pecid RL with source code and dataset used in this paper aiming to maximize the convenience of government staff as well as related researchers to use the tool online.