Autonomous Decision Making on User Complaints with NLP based Deep Learning Models

Abstract

With the advancement in technological domain, the volume of text data, particularly user complaints in public sector, is increasing rapidly. This huge amount of data requires more accurate and automated system for text classification, as manual sorting of this data is time-consuming and error-prone. Addressing this challenge, our research proposes a deep learning-based text classification system that utilizes three models; Convolutional Neural Networks (CNN), a hybrid of Bidirectional Long Short-Term Memory (BiLSTM) and CNN, and the state-of-the-art Bidirectional Encoder Representations from Transformers (BERT) model. These models will interpret the context of text and classify complaints automatically into respective departments. After preprocessing, GloVe embeddings and BERT Tokenizer are used to extract crucial information as features from the complaints, and feed to the models for further pattern learning and classification. Proposed models achieved over 80% classification accuracy on the local complaint dataset, and more than 86% on the internationally acclaimed Consumer Complaint Dataset. Among the models, BERT model outperformed, delivering better performance with accuracy of 82% and 89.5% on the Local and Consumer Complaint Datasets respectively. The proposed methodology will significantly improve complaint management efficiency and serve as a foundation for future improvements in automated text classification.