Tag
The paper introduces evidence-aligned local composition, a method for restoring corrupted documents by inferring soft weightings over frozen domain experts from the marginal evidence of the corrupted observation, achieving high accuracy in tracking expert regions without labeled data.