Prompt-enhanced Federated Learning for Aspect-Based Sentiment Analysis

Aspect-Based Sentiment Analysis (ABSA) involves fine-grained sentiment classification at the entity level, where sentences may contain multiple aspects. Despite notable progress in ABSA with the advent of deep learning and pre-trained language models, data privacy concerns often impede the sharing of sensitive research data. Federated learning (FL) has emerged as a solution, enabling decentralized model training without direct access to raw datasets. However, traditional FL struggles with text data heterogeneity, particularly in ABSA tasks reliant on domain-specific information. This paper presents Prompt-enhanced Federated Learning (PFL) for Aspect-Based Sentiment Analysis. PFL combines FL's privacy-preserving capabilities with prompt tuning. Our model encrypts data using FL to ensure data privacy and exploits external linguistic features for better data representation. Additionally, PFL augments the precision of sentiment analysis by leveraging advanced techniques such as Bidirectional Encoder Representations from Transformer (BERT) for contextual comprehension and Graph Convolutional Network (GCN) for graphical representation, named PFL-GCN. Comprehensive experiments on three benchmark datasets demonstrate PFL-GCN's competitive performance compared to state-of-the-art models while preserving data privacy and enhancing results with prompt tuning. This approach advances sentiment analysis and effectively addresses critical data privacy and security concerns in FL, rendering it well-suited for practical applications that necessitate the rigorous safeguarding of sensitive data.