Title: GlossAssist: Streamlining Corpus Development and Analyzing NLP Model Performance in Low-Resource Documentation Contexts

Abstract: Interlinear glossed text (IGT) serves as the conventional standard for linguistic annotation within the field of language documentation. However, the manual generation of IGT is frequently characterized by high costs and significant time consumption. While automated glossing systems have seen considerable advancements in recent years, their uptake among field linguists remains low. This limited adoption stems largely from the fact that current tools are optimized for evaluation metrics rather than practical utility; they fail to provide an interpretable mechanism for corrections or to integrate linguistic expertise back into the model’s decision-making process.

In response, we introduce GlossAssist, a glossing interface grounded in the retrieval-based architecture of CWoMP (Contrastive Word-Morpheme Pre-training). This design anchors predictions within a dynamic lexicon of learned morpheme representations. When paired with CWoMP, the system operates within an active learning framework: every correction made by an annotator contributes to expanding the lexicon and refining future predictions, thereby enhancing accuracy without necessitating model retraining. This paper details our interface design and posits that such a feedback loop must be considered a fundamental design requirement for NLP tools intended for use by documentary linguists.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC