Autonomous Classification and Decision-Making Support of Citizen E-Petitions Based on Bi-LSTM-CNN

Abstract

The increasing number of e-petition services requires accurate calculation methods to perform rapid and automated delivery. Automated text classification significantly reduces the burden of manual sorting, improving service efficiency. Moreover, existing text classification methods focus on improving sole models with an insufficient exploration of hybrid models. Moreover, existing research lacks combinatorial model selection schemes that yield satisfactory performance for petition classification applications. To address these issues, we propose a hybrid deep-learning classification model that can accurately classify the responsible department of a petition. First, e-petitions were collected from the Chinese bulletin board system and then cleaned, segmented, and tokenized into a sequence of words. Second, we employed the word2vec model to pretrain an embedding matrix based on the e-petition corpus. An embedding matrix maps words into vectors. Finally, a hybridized classifier based on convolutional neural networks (CNN) and bidirectional long short-term memory (Bi-LSTM) is proposed to extract features from the title and body of the petition. Compared with baseline models such as CNN, Bi-LSTM, and Bi-LSTM-CNN, the weighted F1 score of the proposed model is improved by 5.82%, 4.31%, and 1.58%, respectively. Furthermore, the proposed automated petition classification decision support system is available on the e-petition website and can be used to accurately deliver petitions and conduct citizen opinion analysis.