Arabic-emotions-ekman-balanced

Fine-tuned AM5 Arabic Emotion Classification Model

This model is a fine-tuned version of A5 (AnasAlokla/multilingual_go_emotions) for Arabic emotion classification. The model was fine-tuned on the 80% training split of the ArEkmanPlus-balanced dataset, a unified Arabic emotion dataset constructed from multiple existing Arabic emotion resources.

Base Model

• Base model: kiroloskhela/Sentiment-Bert
• Model identifier in the paper: A5
• Task: Arabic text emotion classification
• Training framework: Hugging Face Transformers

Dataset: ArEkmanPlus-balanced

ArEkmanPlus-balanced contains 19,009 Arabic texts collected by unifying several existing Arabic emotion datasets. The dataset focuses on the six basic emotions proposed by Ekman, extended with a None category:

• Anger
• Fear
• Joy/Happiness
• Sadness
• Surprise
• Disgust
• None

The unified dataset was constructed by filtering multi-label instances, unifying equivalent emotion labels, removing duplicate texts and conflicting annotations, and retaining the six Ekman emotions plus the None category.

For this model, 80% of the dataset was used for fine-tuning and 20% was held out for evaluation.

Model Training

The model was fine-tuned as a multi-class sequence classification model. Texts were tokenized using the tokenizer associated with the base model, and the model was trained using the Hugging Face Trainer framework.

The model was evaluated using accuracy and macro F1.

Results

On the held-out test portion of ArEkmanPlus-balanced, the fine-tuned AM5 model achieved:

The model achieved strong performance across the supported emotion categories, with particularly high accuracy for Anger and Fear.

Important: Supported Labels

The original model supports five of the seven ArEkmanPlus labels:

• Anger
• Fear
• Joy/Happiness
• Sadness
• Surprise

The Disgust and None categories are not supported by the original AM5 classification setup used in the experiments.

Therefore, this model should be used only for the five supported emotion categories.

Intended Use

This model is intended primarily for research and experimentation involving emotion classification in Arabic text. It may be useful for computational linguistics, Arabic NLP, emotion analysis, and related research applications.

Limitations

Performance may vary across domains, genres, dialects, and sources of Arabic text. ArEkmanPlus-balanced combines several heterogeneous Arabic emotion datasets, and the original datasets differ in their linguistic characteristics and annotation schemes. The balanced dataset is a sample of the unified dataset with the same number of texts per each emotion class.

The reported performance is based on a single 80/20 train-test split and should not be interpreted as a universal estimate of performance on unseen Arabic text.

Citation

If you use this model or the ArEkmanPlus-balanced dataset in your research, please cite the corresponding paper: TBD