III: Small: A New Machine Learning Approach for Improved Entity Identification

University Of Wisconsin-Madison, Madison WI

Investigators

Abstract

Modern analytics rely on data integration to combine heterogeneous data into a unified repository they can tap into for insights, services, and scientific knowledge. The typical goal of data integration is to combine heterogeneous data about the same real-world entity into a canonical representation of that entity. Traditionally, entity canonicalization methods focus on structured data and leverage the semantics of the schema accompanying the data to come up with canonical entity representations. This dependency on data semantics makes existing entity canonicalization methods inapplicable to dark data, i.e., operational data that corresponds to unstructured, noisy, and incomplete data. This project will develop entity canonicalization methods that focus on unstructured and semi-structured data and are suitable for large-scale integration applications. This work will help ease the currently challenging procedure of heuristically consolidating matching information about the same entity into unified representations and thus enable dark data to be more effectively used in downstream analytics applications. The emphasis of this work is on entity canonicalization techniques that leverage representation learning (a.k.a. feature learning) and deep learning. The combination of distributed representations with deep architectures has emerged as the de facto standard for analyzing and processing unstructured data. This project will develop new deep learning architectures for: (1) record linkage, i.e., clustering unstructured data records that provide information about the same entity; and (2) data fusion, i.e., combining matching unstructured records into a canonical representation of the underlying entity. For record linkage, this work will introduce new deep learning techniques that capture multi-context domain-specific knowledge to learn the semantic similarity between records. For data fusion, this project will design new multi-sequence to one-sequence encoder-decoder recurrent neural networks for data fusion with a particular focus on incomplete data. The outcomes of this project have the potential to advance the state-of-the-art in large scale data integration methods as well as machine learning methods for high-dimensional, sparse, and noisy data. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.