This repository provides implementations for various state-of-the-art models for Emotion Recognition in Conversations (ERC) and Emotion Cause Extraction (ECE). It targets researchers and developers in natural language processing and affective computing, offering a comprehensive toolkit to benchmark and advance conversational AI capabilities.

How It Works

The project implements diverse architectures, including RNN-based (DialogueRNN), Graph Neural Network-based (DialogueGCN), and Transformer-based (COSMIC) models. These methods leverage conversational context, speaker states, and inter-speaker dependencies to predict emotions. COSMIC, in particular, incorporates commonsense knowledge to address challenges like context propagation and emotion shift detection, achieving new state-of-the-art results on several benchmark datasets.

Quick Start & Requirements

• Installation: Environment setup via conda env create -f environment.yml and activation.
• Dependencies: Python 3, PyTorch (versions vary by model, e.g., 1.0+), PyTorch Geometric, Pandas, Scikit-Learn. Some models require TensorFlow and specific CUDA versions for optimal performance.
• Data: Datasets (IEMOCAP, DailyDialog, MELD, EmoryNLP) need to be downloaded separately and placed in specified directories. Pre-trained features (RoBERTa, COMET) are also required for some models.
• Execution: Model-specific Python scripts (e.g., train_iemocap.py, train_dailydialog.py) are used for training and evaluation.
• Resources: Requires significant disk space for datasets and features, and computational resources (GPU recommended) for training.

Highlighted Details

• Implements 10+ distinct ERC and ECE models, including DialogueRNN, DialogueGCN, COSMIC, ICON, and CMN.
• COSMIC achieves state-of-the-art results by incorporating commonsense knowledge for improved context propagation and emotion shift detection.
• Provides implementations for both utterance-level and conversation-level emotion recognition, as well as emotion cause extraction.
• Includes comprehensive performance comparisons across multiple datasets (IEMOCAP, DailyDialog, MELD, EmoryNLP) and metrics (e.g., Weighted F1).

Maintenance & Community

The repository is associated with the declare-lab, with contributions from researchers at various institutions. Recent updates include new work on LLM IQ testing and datasets for humor and emotion cause recognition. A comprehensive reading list for ERC is also available.

Licensing & Compatibility

The repository's licensing is not explicitly stated in the README. However, the inclusion of various models and datasets suggests a focus on academic research. Commercial use compatibility would require further investigation into the specific licenses of each included model and dataset.

Limitations & Caveats

The README indicates that PyTorch Geometric can be a source of non-determinism, and specific execution commands are recommended for result reproducibility. Some models have older dependency requirements (e.g., Python 3.6.5, PyTorch 1.0, TensorFlow 1.9.0), which may pose compatibility challenges with modern environments.