The model is a multi-label classifier over 28 emotional classes for a context-aware scenario. It takes as input a context concatenated with a character description and an utterance, and extracts emotions only from the utterance.

How to use

Here is how to use this model:

python

>>> import torch
>>> from transformers import pipeline
>>> model_name = "roberta-large-emopillars-contextual"
>>> threshold = 0.5
>>> emotions = [
>>>     "admiration", "amusement", "anger", "annoyance", "approval", "caring", "confusion",
>>>     "curiosity", "desire", "disappointment", "disapproval", "disgust", "embarrassment",
>>>     "excitement", "fear", "gratitude", "grief", "joy", "love", "nervousness", "optimism",
>>>     "pride", "realization", "relief", "remorse", "sadness", "surprise", "neutral"
>>> ]
>>> label_to_emotion = dict(zip(list(range(len(emotions))), emotions))
>>> device = torch.device("cuda" if torch.cuda.is_available() else "CPU")
>>> pipe = pipeline("text-classification", model=model_name, truncation=True,
>>>                  return_all_scores=True, device=-1 if device.type=="cpu" else 0)
>>> # input in a format f"{context} {character}: \"{utterance}\""
>>> utterances_in_contexts = [
>>>     "A user watched a video of a musical performance on YouTube. This user expresses an opinion and thoughts. User: \"Ok is it just me or is anyone else getting goosebumps too???\"",
>>>     "User: \"Sorry\", Conversational agent: \"Sorry for what??\", User: \"Don’t know what to do\""
>>> ]
>>> outcome = pipe(utterances_in_contexts)
>>> dominant_classes = [
>>>     [prediction for prediction in example if prediction['score'] >= threshold]
>>>     for example in outcome
>>> ]
>>> for example in dominant_classes:
>>>     print(", ".join([
>>>         "%s: %.2lf" % (label_to_emotion[int(prediction['label'])], prediction['score']) 
>>>         for prediction in sorted(example, key=lambda x: x['score'], reverse=True)
>>>     ]))
surprise: 0.99, amusement: 0.87, curiosity: 0.60, nervousness: 0.58
confusion: 0.97, nervousness: 0.76, embarrassment: 0.65

Training data

The training data consists of 93,979 samples of EmoPillars' context-full subset created using Mistral within our data synthesis pipeline EmoPillars on GitHub. WikiPlots was used as a seed corpus.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 2e-05
train_batch_size: 32
eval_batch_size: 8
seed: 752
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
num_epochs: 10.0

Framework versions

Transformers 4.45.0.dev0
Pytorch 2.4.0a0+gite3b9b71
Datasets 2.21.0
Tokenizers 0.19.1

Evaluation

Scores for the evaluation on the EmoPillars' "context-full" test split:

class	precision	recall	f1-score	support
admiration	0.72	0.68	0.70	635
amusement	0.79	0.63	0.70	211
anger	0.86	0.82	0.84	1155
annoyance	0.80	0.76	0.78	865
approval	0.58	0.42	0.49	250
caring	0.66	0.60	0.63	485
confusion	0.76	0.78	0.77	1283
curiosity	0.83	0.79	0.81	780
desire	0.80	0.75	0.77	864
disappointment	0.79	0.80	0.80	1264
disapproval	0.55	0.47	0.51	445
disgust	0.73	0.60	0.66	320
embarrassment	0.65	0.50	0.57	116
excitement	0.74	0.71	0.73	685
fear	0.87	0.85	0.86	990
gratitude	0.79	0.74	0.76	155
grief	0.79	0.71	0.75	133
joy	0.80	0.78	0.79	668
love	0.70	0.61	0.65	254
nervousness	0.81	0.80	0.80	1368
optimism	0.82	0.76	0.79	506
pride	0.85	0.82	0.83	497
realization	0.74	0.57	0.64	120
relief	0.76	0.67	0.71	211
remorse	0.59	0.53	0.56	206
sadness	0.80	0.79	0.79	922
surprise	0.80	0.78	0.79	852
neutral	0.67	0.57	0.61	392
micro avg	0.78	0.74	0.76	16632
macro avg	0.75	0.69	0.72	16632
weighted avg	0.78	0.74	0.76	16632
samples avg	0.79	0.76	0.75	16632

When fine-tuned on downstream tasks, this model achieves the following results:

task	precision	recall	f1-score
EmoContext (dev)	0.81	0.83	0.82
EmoContext (test)	0.76	0.78	0.77

For more details on the evaluation, please visit our GitHub repository or paper.

Citation information

If you use this model, please cite our paper:

bibtex

@misc{shvets2025emopillarsknowledgedistillation,
      title={Emo Pillars: Knowledge Distillation to Support Fine-Grained Context-Aware and Context-Less Emotion Classification}, 
      author={Alexander Shvets},
      year={2025},
      eprint={2504.16856},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2504.16856}
}

Disclaimer

Click to expand

The model published in this repository is intended for a generalist purpose and is available to third parties. This model may have bias and/or any other undesirable distortions.

When third parties deploy or provide systems and/or services to other parties using this model (or using systems based on this model) or become users of the model, they should note that it is their responsibility to mitigate the risks arising from its use and, in any event, to comply with applicable regulations, including regulations regarding the use of Artificial Intelligence.

In no event shall the creator of the model be liable for any results arising from the use made by third parties of this model.

alex-shvets/roberta-large-emopillars-contextual

roberta-large-emopillars-contextual

Model description