Text-based entity matching for entity resolution and data fusion applied to person descriptions

Text-based entity matching facilitates interoperability between heterogeneous systems by aligning textual person descriptions. We propose an entity matching methodology that integrates rule-based feature extraction, similarity measures, and supervised machine learning classifiers, rigorously evaluated on a person matching problem. We constructed a feature space by extracting domain-specific person attributes from text via a combination of string similarity scores and similarities of inverse document frequency (TF-IDF) embeddings. Next, we evaluated multiple supervised classification models including Multi-Layer Perceptron, Random Forest, and XGBoost, to determine their effectiveness. For evaluation, we created a new domain-specific entity matching dataset named Real Scenario Text-based Person Matching (RSTPM), and assessed the person matching performance of all models in terms of classification metrics and computational cost. In addition, we studied the classification impact of the various features. The proposed approach was shown to achieve an increase of 27.47 percentage points (from 55.41\% to 82.88\%) in F1-Score compared to the baseline and a total Accuracy of 92.14\%, thus demonstrating significant improvements in textual person matching whilst exhibiting a moderate increase in computational demand.