Cross-Slavic Transfer Analysis Framework

1. Core Experimental Setup

Three main conditions:

• Zero-shot transfer: Model trained on Russian/Ukrainian, tested on Belarusian (no Belarusian training data)
• Few-shot transfer: Model trained on Russian/Ukrainian + small Belarusian sample (50/100/500 examples)
• Full supervised: Model trained entirely on Belarusian

Models to test:

• Multilingual models: XLM-R, mDeBERTa-v3, mBERT
• Language-specific: Russian BERT, Ukrainian BERT, Belarusian HPLT BERT
• LLMs: Gemma 2, Llama 3, multilingual models with prompting

2. Transfer Directions to Compare

Russian → Belarusian Ukrainian → Belarusian Polish → Belarusian (bonus: West Slavic baseline) Multi-Slavic (Ru+Uk+Pl) → Belarusian Belarusian only (upper bound)

Key research questions:

• Does Ukrainian transfer better than Russian? (Geographically/politically closer but Russian more prevalent)
• How much Belarusian data is needed to match Russian-only performance?
• Do different tasks show different transfer patterns?

3. Task-Specific Transfer Analysis

For each task (CB, COPA, BoolQ), measure:

A) Performance gaps:

Transfer deficit = Belarusian-only score - Transfer score