---
license: cc-by-sa-4.0
---

# Greek-Lesbian Morphosyntactic Model (Stanza + Greek BERT)

This repository hosts a morphosyntactic model trained using [Stanza](https://stanfordnlp.github.io/stanza/) and fine-tuned with [Greek BERT](https://huggingface.co/nlpaueb/bert-base-greek-uncased-v1) for the **Lesbian dialect of Greek** (spoken on the island of Lesbos). The model has been trained and evaluated on a small, curated treebank of 540 sentences (500 for training, 30 for testing, 10 for development).

The model aims to support part-of-speech tagging, morphological analysis, and dependency parsing for dialectal Greek and is part of a broader effort to document and process regional language varieties.

## 📚 Dataset

The treebank is a manually annotated resource compiled from both **oral** and **written** sources. Oral data were collected between 2023 and 2024 from speakers in various villages of Lesbos:

* **Agra** (Male speaker)
* **Chidira** (Female speaker)
* **Eressos** (Male speaker)
* **Pterounta** (Female speaker)
* **Mesotopos** (Male speaker)
* **Parakoila** (Female speaker)

Written sources include:

* Papanis, D. & Papanis, G. D. (2004). *Lexiko tou Agiasotikou Glosikou Idiomatos*
* Tsokarou-Mitsioni, E. (1998). *Palies Istories ap' tn Agiasiou*
* Tsokarou-Mitsioni, E. (2019). *Prosfygiá*
* Anagnostopoulou, M. A. (2021). *Thematiko Lexiko tis Lesviakis Dialektou*
* Anagnostou, V. T. (2014). *Tsi sta th'ka mas: Komodia sta k'stariot'ka*

The full treebank is openly available here:
🔗 [UD\_Greek-Lesbian on GitHub](https://github.com/UniversalDependencies/UD_Greek-Lesbian)

## 🧠 Model Architecture

* **Base pipeline**: Stanza (v1.7.0+)
* **Pretrained LM**: [Greek BERT](https://huggingface.co/nlpaueb/bert-base-greek-uncased-v1)
* **Tasks**: Tokenization, Lemmatization, POS tagging, Morphological features, Dependency parsing
* **Fine-tuning**: Conducted on the UD\_Greek-Lesbian treebank

## 📈 Performance

Due to the limited size of the training data, the model should be considered **experimental**. It is optimized for research purposes and performs best on dialectal content similar to the training sources. Further fine-tuning and larger datasets will be necessary for production use.

## 📄 Citation

If you use this model or the accompanying treebank, please cite:

> Bompolas, S., Markantonatou, S., Ralli, A., & Anastasopoulos, A. (2025). *Crossing Dialectal Boundaries: Building a Treebank for the Dialect of Lesbos through Knowledge Transfer from Standard Modern Greek*. In Proceedings of the 8th Universal Dependencies Workshop (UDW, SyntaxFest 2025). Association for Computational Linguistics.

## 🔗 Related Resources

* [Universal Dependencies (UD)](https://universaldependencies.org/)
* [Stanza Documentation](https://stanfordnlp.github.io/stanza/)
* [Greek BERT on Hugging Face](https://huggingface.co/nlpaueb/bert-base-greek-uncased-v1)