RoBERTa-EmoWOZ

EmoWOZ Emotion Classifier (RoBERTa)

Fine-tuned RoBERTa-base for emotion classification in dialogue systems.

Model Details

• Base Model: roberta-base (125M parameters)
• Task: 7-class emotion classification
• Training Data: EmoWOZ + DialMAGE (66k+ dialogue utterances)
• Loss Function: Focal Loss (γ=2.0) for class imbalance

Performance (Test Set)

Per-Class F1

• neutral: 0.8882 (6014 samples)
• satisfied: 0.8277 (1817 samples)
• dissatisfied: 0.6931 (604 samples)
• apologetic: 0.7089 (73 samples)
• abusive: 0.7857 (17 samples)
• excited: 0.4310 (91 samples)
• fearful: 0.3500 (18 samples)

Usage

Please download or copy inference.py in this repo to run this model.

Citation

Citations for the original EmoWOZ dataset and this fine-tune model are linked below:

@inproceedings{feng-etal-2022-emowoz,
    title = "{E}mo{WOZ}: A Large-Scale Corpus and Labelling Scheme for Emotion Recognition in Task-Oriented Dialogue Systems",
    author = "Feng, Shutong and Lubis, Nurul and Geishauser, Christian and Lin, Hsien-chin and Heck, Michael and van Niekerk, Carel and Gasic, Milica",
    booktitle = "Proceedings of the Thirteenth Language Resources and Evaluation Conference",
    year = "2022",
    url = "https://aclanthology.org/2022.lrec-1.436",
}

@misc{emowoz-roberta-emotion,
    title = {EmoWOZ RoBERTa Emotion Classifier},
    author = {joshthoo},
    year = {2026},
    publisher = {Hugging Face},
    howpublished = {\url{https://huggingface.co/joshthoo/RoBERTa-EmoWOZ}},
}