Inference-Time Diagnostic for Low-Resource Automatic Post-Editing
September 25, 2026
A black-box diagnostic distinguishes between insufficient training data and model incapacity in low-resource automatic post-editing by varying edit-distance penalties. Testing on English-Sinhala shows the method uses TER-vs-lambda curves and confidence-to-quality correlations to identify failure modes without retraining.
HOW THIS AFFECTS YOU
●
builderYou can use this inference-time signal to evaluate MT quality in low-resource language pipelines.
●
researcherYou can diagnose whether low-resource MT failures stem from data inconsistency or model limits without expensive retraining.