Customer Satisfaction Prediction Using Sentiment Analysis With Bert And Gans

Abstract

The act of predicting customer satisfaction represents a critical aspect for companies that wish to enhance their customer service, customer retention, and decision-making capabilities through a data-driven approach. The conventional approaches to predicting customer satisfaction have been using surveys and classical machine learning models, which cannot be scaled up to accommodate contextual information. The proposed research suggests using the hybrid architecture, which consists of BERT (Bidirectional Encoder Representations from Transformers) and Generative Adversarial Networks (GANs) for reliable sentiment classification and customer satisfaction prediction. Specifically, BERT can be used to obtain contextual embeddings which would take into account nuances in semantic relationships between words, while GANs help overcome class imbalance and data sparsity by generating synthetic samples. The suggested framework is tested on the dataset consisting of 50,000 customer reviews collected from various e-commerce websites. Textual information is preprocessed with regard to tokenization, cleaning, and normalization. The performance of the developed system is estimated via accuracy, precision, recall, and F1-score measures. The conducted research proved the effectiveness of the BERT-GAN hybrid framework, which showed the following results: 91.6% accuracy, 89.8% precision, 88.4% recall, and 89.1% F1-score. This paper illustrates that the incorporation of embeddings from transformers with GAN-created data improves the generalization capabilities of the model, especially for sentiments that are less frequent. This solution provides an organization with an automated framework for processing massive amounts of feedback data. Further research could investigate the adaptation of the model to multiple languages, multimodal feedback, and more complex GAN architectures.