💬 NLP

Transfer Learning for an Endangered Slavic Variety: Dependency Parsing in Pomak Across Contact-Shaped Dialects

This paper addresses dependency parsing for the endangered Pomak language by evaluating zero-shot transfer from a Greek-variety treebank to the Turkish variety and demonstrating that combining cross-variety transfer learning with a new, small manually annotated Turkish-variety corpus significantly improves parsing accuracy.

Original authors: Sercan Karakaş

Published 2026-03-31

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sercan Karakaş

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a master chef who is an expert at cooking Greek-style Pomak soup. This chef has studied thousands of recipes from villages in Greece and can cook that specific soup perfectly.

Now, you ask this chef to cook the Turkish-style Pomak soup for a family in Turkey. The ingredients are similar (both are Slavic languages), but the Turkish version has been cooking next to Turkish and Greek neighbors for centuries. As a result, the Turkish soup uses slightly different spices, adds a pinch of Turkish salt, and sometimes skips the traditional Greek garnish.

When the Greek-trained chef tries to make the Turkish soup, it tastes "okay," but it's not quite right. It's missing the specific local flavor. This is exactly the problem computer scientists faced when trying to teach computers to understand Pomak, a rare and endangered language spoken in Bulgaria, Greece, and Turkey.

Here is a simple breakdown of what this paper did:

1. The Problem: One Recipe Doesn't Fit All

The researchers had a huge digital cookbook (a database) for the Greek variety of Pomak. They wanted to use this to teach a computer how to "parse" (understand the grammar and sentence structure) of the Turkish variety of Pomak.

The Analogy: Think of the computer as a student who only studied the Greek textbook. When the student walks into a Turkish classroom, they get confused because the teacher uses different hand signals and slightly different words to say the same things.
The Result: When the computer tried to understand Turkish Pomak sentences using only the Greek data, it got about 50-59% right. That's like getting a D on a test. It wasn't terrible, but it wasn't good enough to be useful.

2. The First Attempt: A Tiny New Cookbook

The researchers realized they needed a cookbook specifically for the Turkish variety. They gathered native speakers and wrote down 650 sentences in the Turkish dialect.

The Analogy: They tried to teach the student only using this tiny, new 650-page booklet.
The Result: This actually made things worse (dropping to 50% accuracy). Why? Because 650 sentences isn't enough to teach a computer everything about grammar. It's like trying to learn how to drive a car by only reading one chapter of a manual. The student knows the local rules but doesn't understand the general principles of driving.

3. The Solution: The "Master Class" + "Local Intern"

This is where the magic happened. The researchers used a technique called Transfer Learning.

The Analogy: Imagine the student first takes a Master Class using the huge Greek textbook to learn the fundamentals of cooking (how heat works, how flavors balance, the basics of soup-making). Then, they take a short Internship with the Turkish family to learn the local secrets (the specific spices and shortcuts).
The Process:
1. Phase 1 (The Master Class): The computer learns from the big Greek dataset. It learns the general structure of Pomak sentences.
2. Phase 2 (The Internship): The computer is then "fine-tuned" on the small Turkish dataset. It adjusts its brain to recognize the specific Turkish quirks.
The Result: This combination was a huge success! The accuracy jumped to 68%. The computer could now understand the Turkish variety much better because it had the broad knowledge of the Greek data plus the specific local knowledge of the Turkish data.

Why Does This Matter?

Pomak is an endangered language. Very few people speak it, and there are very few written records. Usually, when a language is this small, we can't build good computer tools for it because we don't have enough data.

This paper proves a very important rule for saving endangered languages:

Don't try to build a tool from scratch using only the tiny amount of data you have.

Instead, use the data from a closely related dialect (the Greek variety) to teach the computer the basics, and then use your tiny local data just to fine-tune the tool.

The Big Takeaway

If you want to teach a computer a rare language that has many different "accents" or dialects:

Don't ignore the big cousins: Use the data from the larger, related dialects to build a strong foundation.
Don't ignore the local flavor: Use your small, local data to tweak the model so it sounds authentic.
Mix them together: The best results come from combining the "big picture" with the "local details."

This approach gives hope that we can build useful technology for many other endangered languages, even when we only have a handful of sentences to work with.

1. Problem Statement

The paper addresses the challenge of Natural Language Processing (NLP) for Pomak, an endangered Eastern South Slavic language spoken in Bulgaria, Greece, and Turkey. The core problem is twofold:

Low-Resource Constraints: Pomak lacks a widely adopted orthographic standard and has very limited annotated data.
Dialectal Heterogeneity & Contact Effects: Pomak exists as a dialect continuum heavily shaped by contact with Greek and Turkish. The study focuses on the Turkey variety (Uzunköprü) versus the Greece variety (Xanthi/Paşevik).
- These varieties exhibit systematic grammatical differences (morphosyntax, case marking, argument realization) due to contact-induced change, not just lexical variation.
- The Specific Question: Can a dependency parser trained on the existing Universal Dependencies (UD) treebank (primarily the Greece variety) effectively generalize to the Turkey variety? If not, how can limited data from the Turkey variety be leveraged to improve performance?

2. Methodology

A. Data Resources

The study utilizes and creates three distinct datasets:

Dataset A (Source): The existing Pomak UD treebank (Jusúf Karahóğa et al., 2022), primarily representing the Greece variety.
Dataset B (Target Mini-Corpus): A newly compiled corpus of 650 sentences representing the Turkey variety (Uzunköprü).
- Sources: Six native bilingual speakers (Pomak/Turkish) and prior descriptive linguistic data.
- Annotation: Manually annotated following the UD framework (POS, morphology, dependencies).
- Split: 520 sentences for training, 130 for development.
Dataset C (Target Test Set): A held-out test set of 90 sentences from a different native speaker than those in Dataset B to ensure speaker-disjoint evaluation.

B. Model Architecture

The authors implement a state-of-the-art biaffine, graph-based neural dependency parser (based on Dozat et al., 2017).

Input: Token representations combining word embeddings, character-level embeddings (via CNN/CharEnc), and discrete features.
Contextualization: A BiLSTM encoder generates contextual representations.
Scoring:
- Arc Scoring: A biaffine scorer predicts the probability of a head-dependent relationship.
- Label Scoring: A separate biaffine scorer predicts the dependency relation label.
Decoding: Uses the Chu-Liu/Edmonds algorithm to find the maximum spanning arborescence (non-projective allowed).
Loss Function: Negative log-likelihood of head selection and relation labeling, plus L2 regularization.

C. Experimental Design

The authors compare three training regimes to evaluate cross-variety transfer and adaptation:

Baseline (Zero-Shot Transfer): Train on Dataset A (Greece) $\rightarrow$ Test on Dataset C (Turkey).
In-Variety (Small Data): Train exclusively on Dataset B (Turkey mini-corpus) $\rightarrow$ Test on Dataset C.
Transfer/Adaptation: Train on Dataset A (Greece) $\rightarrow$ Fine-tune on Dataset B (Turkey) $\rightarrow$ Test on Dataset C.

3. Key Contributions

New Resource: Creation of the first manually annotated, UD-compatible corpus for the Turkey variety of Pomak (650 sentences), addressing a critical gap in low-resource NLP for this dialect.
Empirical Analysis of Contact-Induced Variation: Quantitative evidence demonstrating that dialectal differences in Pomak are not merely lexical but involve systematic morphosyntactic shifts (e.g., dative marking strategies like na vs. inflectional case) that degrade parser performance.
Methodological Framework for Endangered Languages: A demonstration that for endangered languages with small dialect-specific corpora, transfer learning with fine-tuning is superior to training from scratch or relying solely on out-of-variety data.

4. Results

The performance is measured using Unlabeled Attachment Score (UAS) and Labeled Attachment Score (LAS).

Training Regime	UAS	LAS	Analysis
Baseline (A $\to$ C)	59%	51%	Zero-shot transfer from Greece to Turkey shows moderate performance but a significant gap (8 points) between UAS and LAS, indicating that relation labeling is particularly sensitive to dialectal differences.
In-Variety (B $\to$ C)	50%	44%	Training only on the small Turkey corpus (650 sentences) performs worse than the baseline. This highlights the "extreme low-resource" bottleneck: 650 sentences are insufficient for the model to learn robust syntactic generalizations, even if the data is dialect-matched.
Transfer/Adaptation (A $\to$ B $\to$ C)	68%	62%	Best Performance. Combining the broad coverage of the Greece UD treebank with fine-tuning on the Turkey mini-corpus yields the highest scores. This represents a +9 UAS / +11 LAS gain over the baseline and a +18 UAS / +18 LAS gain over the in-variety-only model.

5. Significance and Discussion

Dialect Sensitivity is Crucial: The results confirm that Pomak varieties are not interchangeable for NLP models. The "contact-shaped" nature of the Turkey variety (influenced by Turkish) creates distributional shifts in functional markers and argument realization that standard parsers fail to capture without adaptation.
The "Small Data" Paradox: The study reveals a critical tension in endangered language NLP: while dialect-matched data is theoretically ideal, the extreme scarcity of such data (e.g., 650 sentences) often leads to poor performance due to lack of coverage.
Optimal Strategy: The most effective approach for endangered, dialectally diverse languages is Transfer Learning with Fine-Tuning.
- The larger source corpus (Greece) provides the necessary statistical strength and syntactic coverage.
- The small target corpus (Turkey) acts as an "adaptation signal," re-anchoring the model's parameters to the specific distributional properties (e.g., specific case markers, word order preferences) of the target dialect.
Implications for Resource Design: The authors argue for explicit metadata tagging (dialect, region) in UD treebanks and targeted annotation strategies that prioritize constructions known to vary across dialects (e.g., oblique marking, ditransitives).

In conclusion, the paper demonstrates that while cross-variety transfer in Pomak is imperfect due to contact-induced grammatical divergence, a hybrid strategy leveraging large out-of-variety resources and small in-variety adaptation sets can significantly overcome data scarcity, providing a viable path for building syntactic tools for endangered languages.