AWS has demonstrated a workflow for correcting and harmonizing metadata so datasets with different labels, identifiers and formats can be analyzed together. It combines conventional validation with language models when simple string matching cannot resolve a mismatch.
The system checks whether columns match a target schema and whether individual values satisfy required fields, controlled vocabularies and formatting rules. Embeddings can map semantically similar terms, while models in Amazon Bedrock handle harder cases such as domain-specific synonyms or deciding that one source column should become several target columns. Storage, authentication, job tracking and compute use S3, Cognito, DynamoDB and ECS.
The default workflow returns recommendations to a user for approval, preserving domain expertise and an audit point. AWS also discusses more autonomous agent-driven operation, but that raises the need for confidence thresholds, versioned schemas, monitoring and rollback. Language models can accelerate messy edge cases; they do not remove the requirement to verify scientific meaning before changing shared data.