Technical Summary: GlossAssist

Problem Statement

Interlinear Glossed Text (IGT) is the standard format for linguistic annotation in language documentation, yet its manual creation is slow, costly, and requires extensive morphological analysis. While automated glossing systems have improved in benchmark performance (e.g., segmentation, glossing, and translation across diverse languages), their adoption among field linguists remains low. The paper identifies that this lack of adoption is not solely due to raw accuracy or model hallucinations, but rather a fundamental design mismatch: existing tools are optimized for evaluation rather than practical use. They function as "black boxes" that produce outputs difficult to inspect, correct, or trust, treating the linguist as a passive consumer of predictions rather than an active collaborator. Furthermore, many existing tools are inaccessible due to complex code repositories or lack of maintenance, and they fail to provide a mechanism for incorporating linguistic expertise back into model behavior without full retraining.

Methodology

The authors present GlossAssist, a glossing tool designed to integrate human expertise directly into the model's refinement process. The system is built upon two primary components:

1. Retrieval-Based Architecture (CWoMP): GlossAssist utilizes the Contrastive Word-Morpheme Pre-training (CWoMP) architecture. Unlike sequence-to-sequence models that generate unconstrained text, CWoMP grounds predictions in a mutable lexicon of learned morpheme representations. This design prevents the generation of unattested morpheme types (hallucinations) by design, as predictions are retrieved from a discrete, verified lexicon.
2. Interactive Annotation Interface: The tool features a three-panel interface:
   • Left Panel (Evidence): Displays the source text alongside retrieved lexical entries and grammatical rules relevant to the input. This provides transparent evidence for the model's predictions.
   • Center Panel (Workspace): Presents word-by-word segmentation and glossing predictions. Annotators can accept, reject, or modify these predictions. Crucially, accepted segment/gloss pairs are automatically added to the mutable lexicon via a toggleable mechanism.
   • Right Panel (Translation): Displays model-predicted translations with confidence scores, allowing independent acceptance or modification.

Active Learning Loop: The system treats every annotator correction as part of an active learning setting. When a linguist accepts a prediction, it is absorbed into the lexicon and retrieved for future sentences. This allows the system to expand its coverage and improve future predictions without the need to retrain the underlying model.

Researcher Dashboard: A secondary component aggregates feedback data across annotation sessions, providing session-level Key Performance Indicators (KPIs) such as average time per sentence, prediction accuracy, and total feedback events. This facilitates workflow-centered evaluation.

Key Contributions

• A New Design Paradigm: The paper argues for a shift from minimizing human annotation effort to treating annotation decisions as investments that refine the model. GlossAssist reframes the annotation session where accepts become validated lexicon entries, rejects signal model failures, and modifications serve as gold-standard corrections.
• Interpretability and Trust: By grounding predictions in a retrievable lexicon and displaying the supporting evidence (lexical entries and rules) to the user, the tool addresses the "black box" problem. It ensures that segment-gloss alignment is structurally guaranteed and prevents the hallucination of unattested morpheme types.
• Workflow-Centered Evaluation: The authors propose that evaluation for documentary linguistics should move beyond isolated corpus-level metrics (like MER or WER) to include workflow metrics such as annotation speed, model adaptability, and the ability to diagnose errors. The GlossAssist dashboard is designed to capture the data necessary for this type of evaluation.

Results and Claims

The paper does not present new empirical benchmark results comparing GlossAssist against other systems on standard test sets. Instead, it relies on the following claims and supporting evidence:

• Theoretical and Empirical Support from CWoMP: The authors cite Alper et al. (2026), noting that CWoMP's mutable lexicon evaluation demonstrates that extending the lexicon with ground-truth morphemes consistently reduces Morpheme Error Rate (MER), with the most significant gains occurring in low-resource settings.
• Alignment with User Needs: The design is explicitly motivated by findings from Rice et al. (2025), which identified misaligned segment-gloss pairs and hallucinated morpheme types as the most damaging failure modes for field linguists. GlossAssist is engineered to mitigate these specific issues.
• Efficiency Curve: The paper posits a "compounding efficiency curve": while early annotation sessions require more correction, the growing lexicon improves model coverage over time, reducing the proportion of predictions requiring intervention. This aligns with the natural workflow of field linguists building language descriptions over time.

Significance

The paper argues that the gap between automated glossing research and field adoption is structural, not incidental. By treating the linguist as an expert collaborator and providing a mechanism for immediate, interpretable feedback that updates the system without retraining, GlossAssist aims to bridge the divide between NLP research and documentary practice. The authors position this feedback loop as a necessary design requirement for NLP tools intended for language documentation. Future work involves conducting user studies with practicing documentary linguists to measure annotation efficiency and the effects of active learning across low-resource languages, as well as integrating with existing tools like ELAN and FLEx.