A POS Tier Is the Key to Automated Annotation for Low-Resource Language Documentation: Neural Interlinear Glossing for Irabu, a Southern Ryukyuan Language

Documentary linguists face a severe 'transcription bottleneck,' where the time required to annotate discourse data—often one hour of work per minute of recording—limits the amount of data that can be processed. This study investigates how to optimize this process by asking two questions: what is the minimum amount of annotated data required for effective automation, and does adding a part-of-speech (POS) tier (creating a 'quadrilinear' annotation format) improve the results of an automated interlinear glossing pipeline?

The author implements a full neural annotation pipeline (morpheme segmentation, POS tagging, and glossing) for Irabu, a Southern Ryukyuan language. The models are deliberately small, transparent BiLSTM-CRF architectures, chosen to align with the linguist's reasoning process and to function effectively with the limited data available in endangered language documentation. The study evaluates performance across different training budgets (from 6 to 47 minutes of discourse) and compares trilinear annotation (text, gloss, translation) against quadrilinear annotation (text, POS, gloss, translation).

The study demonstrates that a POS tier provides a clear, measurable benefit to automated glossing. When using 'gold' (perfect) POS tags, grammatical glossing accuracy improves by 4.4 percentage points. This benefit is even more pronounced in low-data regimes, where the gain reaches 11.6 points, effectively cutting the required manual annotation effort in half. However, in a fully automatic pipeline, this gain is currently offset by tagger errors. The author identifies a 'break-even' point at approximately 88% tagger accuracy; above this threshold, the POS tier begins to provide a net positive contribution to the glossing model.

This research provides an empirical basis for documentation workflows that currently rely on intuition or conventional 'hour-count' targets. By quantifying the return on investment for different annotation tiers, the author offers a concrete recommendation: linguists should adopt a quadrilinear annotation format. The study also proposes a two-stage documentation strategy where a small, high-quality corpus is annotated first to train the pipeline, which then assists in the annotation of subsequent, larger volumes of discourse.