Towards a Greek Proverb Atlas: Computational Spatial Exploration and Attribution of Greek Proverbs

John Pavlopoulos, Panos Louridas, Panagiotis FilosEMNLP 2024 · DOI

Introduction

Proverbs are more than just "old wives' tales"; they are "carriers of oral wisdom" that encapsulate cultural and historical identities. While linguists have long studied regional dialects, doing so at scale for proverbs has been difficult. This research addresses two main challenges: the lack of machine-actionable data for Greek proverbs and the difficulty of geographical attribution due to the "regularization" of dialects by historical collectors, which often erases local linguistic fingerprints.

The Greek Proverb Atlas Dataset

The study introduces a massive new resource: the Greek Proverb Atlas Dataset, containing 134,493 proverb variants collected from the Hellenic Folklore Research Centre, dating back to 1807. After cleaning and removing duplicates, the researchers established a balanced benchmark of 11,500 proverbs across 23 distinct locations. This dataset is publicly available to facilitate further research in computational folklore.

Examples of Greek proverbs with transliteration, translation, and linguistic distance metrics.

Methodology

The researchers combined traditional folklore studies with modern Natural Language Processing (NLP) techniques:

  • TFILF (Term Frequency-Inverse Location Frequency): A novel adaptation of TF-IDF designed to identify tokens unique to specific geographical locations.
  • Authorship Analysis (AA): Treating each location as a single "author," the team trained 3-gram character-based language models to determine if a proverb's linguistic style matched a specific region.
  • Benchmark Models: Conventional machine learning models (Logistic Regression, SVM, Random Forests) were compared against GrBERT, a Transformer model pre-trained on contemporary Greek.
  • Conformal Prediction: To handle proverbs lacking location data, conformal prediction was used to generate statistically rigorous "prediction sets" of possible origins rather than a single guess.

Key Insights: Linguistics vs. Semantics

A primary finding is that how a proverb is stated (phonology and syntax) is more indicative of its origin than what is being said (semantics). Character n-gram models often outperformed deep learning models like BERT because they captured dialectal "glitches"—such as tsitakism (palatalization)—that signal specific regions. Interestingly, GrBERT struggled with these regional peculiarities, suggesting that transfer learning from standard modern Greek is not always effective for dialectal or historical data.

Results

The models achieved a macro-averaged F1 score of 0.30, significantly outperforming a random classifier (0.04). Regional performance varied greatly; Cyprus was the most identifiable region with an F1 score of 0.80, followed by Pontos and Skyros. For geocoding tasks, the best models achieved a Mean Absolute Error of approximately 145km for latitude and 163km for longitude.

Confusion matrix for proverb classification across regions, highlighting high accuracy for Cyprus and distinct clusters.

Conclusion and Future Work

While collector bias in the 19th century remains a challenge, this work successfully bridges folklore and computational linguistics. Future research will aim to incorporate temporal data to track the evolution of proverbs over time and explore "linguistic paths" along historical trade routes to understand the economic factors influencing the spread of oral wisdom.

Summary updated

BibTeX
@inproceedings{pavlopoulos2024towards,
  author = {John Pavlopoulos and Panos Louridas and Panagiotis Filos},
  title = {{Towards a Greek Proverb Atlas: Computational Spatial Exploration and Attribution of Greek Proverbs}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2024},
  doi = {10.18653/v1/2024.emnlp-main.661},
}
Figure