LEMON: Explainable Entity Matching

Entity matching systems lack transparency: State-of-the-art AI methods for matching records across databases are "hard to interpret," making it difficult for users to understand why the system made a particular decision—a critical gap for responsible AI deployment in data management.

Standard explainability tools need customization for matching tasks: Popular explanation methods like LIME don't work effectively for entity matching without significant adaptation, highlighting that responsible AI solutions must be tailored to specific applications rather than applied generically.

LEMON improves human understanding substantially: User studies showed that LEMON explanations increased the rate at which people could construct counterfactual examples (understanding what would need to change for a different outcome) from 54% to 64% for matches and from 15% to 49% for non-matches, demonstrating meaningful improvements in interpretability.

The method addresses real-world explanation challenges: LEMON tackles three specific problems—cross-record interactions, explaining why records didn't match, and adjusting explanation detail to match system sensitivity—suggesting that effective explainability requires understanding domain-specific complexities.

Model-agnostic approach increases practical applicability: Because LEMON works with different matching systems without requiring access to internal model details, it can be more readily adopted across organizations using different entity matching tools.