EmoAlign

# EmoAlign: Multi-Grained Empathetic Dialogue System with Personality Fusion **EmoAlign** is a novel dialogue generation framework that enhances emotional support conversations through cognitive affective fusion, memory alignment, and personality-aware planning. It is designed for scenarios requiring empathetic, context-sensitive, and supportive interaction, such as mental health counseling, social chatbots, and AI companions. --- ## 📌 Overview This project implements the core framework proposed in the paper: > **"Affective Fusion for Emotional Support: An Empathetic Dialogue Framework Using Multi-Grained Cognition and Personality Alignment"** **EmoAlign** integrates three main components: - **Perceptual-Cognitive Fusion (PCF):** Fuses affective cues and multi-turn cognitive representations using hierarchical attention. - **Hierarchical Memory Fusion (HMF):** Incorporates latent intent, role transitions, and multi-scale memory to maintain emotional and semantic consistency. - **Causal Persona Planning (CPP):** Models target user personality and alignment trajectory to generate context-consistent and personalized responses. --- ## 🌟 Key Features - 🎯 Multi-grained fusion of emotion, cognition, and intent- 🧠 Long-range dialogue memory encoding- 👤 Target speaker personality alignment- 🧾 Explicit emotion and intent constraints- 📈 State-of-the-art performance on ESConv, EmpatheticDialogues --- ## 🧩 Project Structure ```bash.├── data/ # ESConv / EmpatheticDialogues dataset loaders├── models/ # PCF, HMF, CPP modules├── configs/ # Model and training configuration├── scripts/ # Train / evaluate / inference scripts├── utils/ # Helper functions and loss modules├── results/ # Output results, logs, visualizations├── checkpoints/ # Trained model weights└── README.md