This model is a fine-tuned google-bert/bert-base-multilingual-cased sequence classifier for multilingual emotion recognition in English and Spanish dialogue utterances.

Source code: Mario-RC/multilingual-emotion-classifier

Model Details

Model repository: mario-rc/emotional-classifier-bert-base-multilingual-cased
Base model: google-bert/bert-base-multilingual-cased
Architecture: BertForSequenceClassification
Task: text classification / emotion classification
Languages: English, Spanish
Max sequence length: 128 tokens
Number of labels: 7

Emotion Labels

The model predicts one of seven normalized emotion labels:

anger, disgust, fear, happiness, neutral, sadness, surprise

Training Data

The training pipeline combines DailyDialog and EmpatheticDialogues-derived CSV resources into a multilingual English/Spanish dataset. EmpatheticDialogues labels were mapped into the seven normalized categories above, while ambiguous or underrepresented labels were removed. The training split was resampled to reduce the majority neutral class and upsample minority classes.

Training Setup

Framework: Hugging Face Transformers
Base checkpoint: google-bert/bert-base-multilingual-cased
Task: sequence classification
Max sequence length: 128
Epochs: 3
Learning rate: 5e-6
Batch size: 32
Dropout: 0.2
Seed: 42

Evaluation

The confusion matrix shows true vs. predicted emotion labels on the multilingual test split. Diagonal cells indicate correct classifications, while off-diagonal cells show the main confusions between emotion classes.

Model Comparison

Model	Base model	Test accuracy	Test Macro F1	GPT-4 ES benchmark	GPT-4 EN benchmark
Emotional Classifier BERT-Base-Multilingual-Cased	`google-bert/bert-base-multilingual-cased`	0.5506	0.5468	72.91%	70.81%
Emotional Classifier BERT-Base-Multilingual-Uncased	`google-bert/bert-base-multilingual-uncased`	0.5560	0.55	74.25%	74.16%
Multilingual Emotional Classifier XLM-RoBERTa-Base	`FacebookAI/xlm-roberta-base`	0.5171	0.5078	82.61%	76.17%
Multilingual Emotional Classifier XLM-RoBERTa-Large	`FacebookAI/xlm-roberta-large`	0.6640	0.6658	85.95%	79.19%

Usage

python

from transformers import AutoModelForSequenceClassification, AutoTokenizer, pipeline

model_id = "mario-rc/emotional-classifier-bert-base-multilingual-cased"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)

classifier = pipeline("text-classification", model=model, tokenizer=tokenizer)

print(classifier("I feel great today."))
print(classifier("Estoy preocupado por manana."))

Limitations

The model is designed for short dialogue utterances and seven broad emotion categories. Predictions may be less reliable for long documents, sarcasm, mixed emotions, domain-specific language, or languages beyond English and Spanish.

Citation

This work is detailed in Section 4.4.3, User Emotion Recognition, of:

Personal Assistant with Emotional and Multilingual Capabilities for Social Robots
M. Rodriguez-Cantelar, PhD Dissertation, Universidad Politecnica de Madrid, 2025.