BERT-based Consumer Sentiment Analysis for Personalized Marketing Strategies

Abstract

This study aims to construct an efficient consumer sentiment analysis model and apply it to optimize personalized marketing strategies for enhancing user conversion rates and satisfaction. The yf_amazon dataset and Chinese sentiment classification corpus are selected as foundational data by integrating natural language processing and deep learning technologies. Preprocessing procedures including tokenization, stop word removal, noise filtering, and sentiment annotation, are employed to clean and structure consumer reviews. A sentiment classification model based on Bidirectional Encoder Representations from Transformers (BERT) and Transformer architectures is developed and fine-tuned with parameters including a learning rate of 2e-5 and 4 training epochs. Evaluation metrics encompass accuracy, precision, recall, and F1 score. Experimental results demonstrate that the proposed model achieves 92.5% accuracy and a 0.90 F1 score in sentiment classification tasks. This model outperforms conventional Convolutional Neural Networks (83.6% accuracy, 0.83 F1 score) and Long Short-Term Memory models (86.2% accuracy, 0.85 F1 score). Further segmentation of user sentiment clusters based on analysis results enables the implementation of differentiated recommendation strategies and timelimited marketing campaigns. Ultimately, it increases user conversion rates from 25.0% to 32.7% (a 7.7 percentage point improvement) while significantly enhancing consumer satisfaction with personalized recommendations. These findings indicate that Transformer-based sentiment analysis models hold substantial practical value for advancing personalized marketing applications.