luvnpce83/ancient-greek-emotion-bert

Ancient Greek Emotion BERT

This is a model fine-tuned from pranaydeeps/Ancient-Greek-BERT for 8-class emotion classification on Koine Greek. The model predicts one of eight basic emotions based on Plutchik’s Wheel of Emotions: Joy, Trust, Fear, Surprise, Sadness, Disgust, Anger, and Anticipation.

This model was developed as part of the doctoral dissertation "Emotions in the Identity of Paul: A Sentiment Analysis Approach to Paul’s Jewish Identity" at Yonsei University.

Model Description

This model is designed to perform sentence-level emotion analysis on Koine Greek texts, particularly those from the Pauline Epistles and related literature. It was trained on a custom-built dataset derived from the Louw-Nida lexicon and augmented using modern NLP techniques to overcome the challenges of a low-resource language.

Base Model

The fine-tuning process started from the pranaydeeps/Ancient-Greek-BERT model, a BERT model pre-trained specifically on a large corpus of Ancient Greek texts. Utilizing this specialized base model was crucial for achieving high performance on this downstream task.

Intended Use & Limitations

Intended Use

This model is intended for academic research in Digital Humanities, Biblical Studies, Classics, and Linguistics for analyzing emotional expression in Koine Greek texts. It can be used to generate quantitative data on the emotional valence of sentences for large-scale textual analysis.

Limitations

• Domain Specificity: The model was trained on a dataset primarily derived from the New Testament. Its performance on other genres of Koine Greek (e.g., philosophy, history, poetry) has not been evaluated and may be suboptimal.
• No Absolutes: Emotion is subjective and context-dependent. The model's predictions should be used as a tool for analysis, not as a definitive judgment of a text's emotional content.
• Potential Bias: The model reflects the emotional expressions present in its training data, which is sourced from ancient religious texts. It may carry the biases inherent in that source material.

How to Use

The easiest way to use this model is with a pipeline from the 🤗 Transformers library.

from transformers import pipeline

# Replace "YourUsername/your-model-name" with your actual model ID
classifier = pipeline("text-classification", model="luvnpce83/ancient-greek-emotion-bert")

result = classifier("ὦ ἀνόητοι Γαλάται")
print(result)
# [{'label': 'Anger', 'score': 0.8521...}]

Training Data

The model was trained on a custom dataset of 2,616 annotated Koine Greek sentences. The creation process involved:

1. Initial Curation: A "golden standard" corpus of 884 samples was manually created based on the semantic domains of the Louw-Nida lexicon.
2. Data Augmentation: The dataset was expanded using back-translation and generative augmentation via a large language model to enhance robustness.

Training Procedure

The model was fully fine-tuned from a bert-base-cased checkpoint. Key hyperparameters include:

• Learning Rate: 5e-5
• Batch Size: 64
• Optimizer: AdamW
• Epochs: 12 (with early stopping, patience=5)
• Max Sequence Length: 512

The training was managed and logged using Weights & Biases.

Evaluation Results

The model achieved a Macro F1 score of 0.680 on the held-out validation set.

Citation

If you use this model in your research, please cite the following dissertation:

@phdthesis{kang2025emotions,
  author    = {Kang, Young Un},
  title     = {Emotions in the Identity of Paul: A Sentiment Analysis Approach to Paul’s Jewish Identity},
  school    = {School of Theology, Yonsei University},
  year      = {2025}
}