Cross-Lingual Transfer Efficiency in Low-Resource Morphosyntactic Parsing: A Contrastive Annotation-Theoretic Framework for Dependency Treebank Calibration

Abstract

The calibration of dependency treebanks for low-resource languages remains a critical bottleneck in cross-lingual transfer learning pipelines, particularly when morphosyntactic divergence between source and target languages is pronounced. This study proposes a contrastive annotation-theoretic framework that systematically reconciles Universal Dependencies (UD) schema inconsistencies across typologically distant language pairs. Drawing on a multi-corpus evaluation spanning six agglutinative and fusional languages, we operationalize a weighted inter-annotator agreement metric sensitive to head-attachment ambiguity and functional category misalignment. Results demonstrate a statistically significant improvement in labeled attachment score (LAS) of 6.3–9.1 percentage points over baseline transfer models. The framework further offers reproducible guidelines for schema harmonization, establishing a scalable methodological benchmark applicable to under-documented language documentation initiatives and second language acquisition corpus research.