Data Integration and Explainable Entity Matching

Methods for matching, linking, cleaning and explaining heterogeneous records in data integration pipelines.

Entity Matching (EM) aims to identify records that refer to the same real-world entity across heterogeneous data sources. It is a central task in data integration, data cleaning, deduplication and the construction of reliable data-intensive applications.

DTALab studies Entity Matching not only as a prediction task, but as an interpretable and auditable component of data integration pipelines. The goal is to design methods that can decide whether two records match, explain why they match or do not match, and support users in understanding the evidence behind model decisions.

The research line focuses on three main directions:

  1. Explainable Entity Matching: methods that explain the behaviour of Machine Learning and Deep Learning models for EM, including post-hoc explanation techniques and model-aware diagnostics.

  2. Intrinsically interpretable EM models: approaches that produce matching decisions through meaningful intermediate representations, such as decision units and interpretable feature spaces.

  3. Evaluation and analysis of neural EM models: studies of how BERT-based and Transformer-based models perform Entity Matching, which components drive their predictions, and how their decisions can be inspected.

Recent work in this area includes Landmark Explanation, an explainer designed for Entity Matching models; WYM, an intrinsically interpretable EM system based on decision units; and CREW, a cluster-based explanation method that groups words according to semantic similarity, dataset structure and their importance for the matching decision.

This research contributes to trustworthy data integration pipelines where matching decisions must be accurate, understandable and usable by domain experts.

Selected publications