Fine-Grained Text Sentiment Analysis and Transfer

implementing BERT-based Classification, T5 Transformer, and Representation Engineering

Team Contribution:
Dataset Generation and Annotation, Model Architecture and Development, System Integration and Validation

This project presents a comprehensive emotion processing pipeline that combines three interconnected tasks: emotion classification, intensity estimation, and emotion transfer in text. We implement a BERT-based classifier to categorize text into six distinct emotions (anger, caring, confusion, fear, joy, and sadness), achieving 98.6% validation accuracy. Building upon this, we develop a regression model for quantifying emotion intensity on a normalized scale, followed by two approaches for emotion transfer: representational engineering and T5-based text generation. The representational engineering approach manipulates hidden states to control emotional expression, while the T5 model generates text with specified emotion intensities. Our integrated pipeline demonstrates effective end-to-end emotion processing, enabling both analysis and modification of emotional content in text. Experimental results show successful emotion transformation while maintaining semantic coherence, with potential applications in empathetic AI systems and emotional content generation.

Links Project

See our report here: