declare-lab/conv-emotion

Implements graph neural networks and attention-based RNNs that model speaker states and inter-party dependencies across multiple conversation datasets (IEMOCAP, DailyDialog, MELD, EmoryNLP). The COSMIC model integrates commonsense knowledge graphs with dialogue context to achieve state-of-the-art results, while supporting both PyTorch and TensorFlow implementations. Also includes modules for emotion cause extraction and multimodal feature integration via companion datasets.

1,502 stars. No commits in the last 6 months.