Automated identification and adaptation prediction of loanwords in low-resource languages

Abstract

Automated loanword identification remains a challenge in Natural Language Process- ing (NLP), particularly for low-resource languages under severe lexical pressure. This thesis investigates the automated extraction and morphological classification of lexical borrowings across three typologically distinct languages: Asturian, Basque, and Modern Greek. Our primary objectives are formalizing a morphological taxonomy, constructing training corpora via synthetic data generation (specifically, of the tech domain), and evaluating architectural baselines. We test non-contextual heuristics, distantly super- vised multilingual encoders (XLM-RoBERTa, mmBERT), and generative Large Language Models (Qwen3.5-9B) via zero and few-shot In-Context Learning, assessing their ability to both identify and classify borrowings by assimilation level. Our findings reveal an ar- chitectural dichotomy. Contextual encoders function as high-precision extractors but are obstructed by low recall due to sub-word tokenization fragmentation, consistently operat- ing in a high-precision/low-recall regime regardless of the training data source. Attempts to boost recall via external distant supervision introduce precision loss through cognate confusion and typological transfer barriers. Conversely, generative models exhibit a low-...