Loading...
Source-free Unsupervised Domain Adaptation of NER Tasks in Ancient Greek
Liu, Lilly
Liu, Lilly
Abstract
In this paper, we introduce a Source-Free Unsupervised Domain Adaptation (SFUDA) framework for Named Entity Recognition (NER) in Ancient Greek. Ancient Greek is characterized by limited labeled corpora, high annotation cost, and often substantial domain or language mismatch between available training data and the target setting. Our approach treats the problem as knowledge distillation from a Large Language Model (LLM) teacher to a small discriminative student, requiring no ground truth target labels at any stage. We employ GPT-4o-mini as a zero-shot pseudo-labeler over the NEReus Ancient Greek corpus and introduce an confidence score that combines GPT-4o-mini's log-probabilities and Gaussian Mixture Model clustering of 150-dimensional FastText word embeddings trained on Ancient Greek to measure the reliability of each pseudo-label. These per-token confidence scores are used to weight the cross-entropy loss during finetuning of an XLM-RoBERTa student model, amplifying reliable pseudo-labels and suppressing uncertain ones.
Description
Date
2026-05-01
Journal Title
Journal ISSN
Volume Title
Publisher
Collections
Download Dataset
Rights Holder
The author
Usage License
Embargo
Research Projects
Organizational Units
Journal Issue
Keywords
Citation
Advisor
Department
Computer Science