CRII: III: Knowledge Graph Completion with Transferable Representation Learning

Knowledge graphs (KGs) provide both open-world and domain-specific knowledge representations that are integral to many artificial intelligence (AI) systems. However, constructing KGs is usually very costly and requires extensive human effort. While representation learning offers a solution by automatically inferring missing knowledge in the embedding space, KGs are constructed independently thereby missing complementary knowledge available in other KGs.

In fact, the independently created knowledge is often interrelated across different perspectives. For example, knowledge about The Tale of Genji (the oldest Japanese novel) in an English KG may be enriched with complementary knowledge readily available in a Japanese KG; in the proteomics domain, verification of Protein-protein interaction is also in parallel to the study of the genomic functions of the proteins that are annotated by the Gene ontologies.

In particular, this project studies a novel direction of transferable representation learning for KGs, which seeks to associate the interrelated knowledge from different isolated sources in a common embedding scheme using minimal supervision, and allowing complementary knowledge to easily migrate across different KGs. The outcome of this project will create universal knowledge representation for different sources, fields and languages, therefore supporting applications with cross-domain or low-resource decision making.

The outcome will broadly improve the utility of the knowledge in computational research including AI, natural language understanding (NLU) and recommender systems, as well as interdisciplinary research in biology, pharmacology and social sciences.

The goals of this project are to develop new data-driven machine learning methods for automatic KG completion. Such a method will be able to automatically combine isolated KGs, allow complementary knowledge to transfer across them, and efficiently infer globally missing knowledge based on what is known. This project will systematically solve several key technical challenges of leveraging incidental and auxiliary supervision signals to capture various types of knowledge association.

Furthermore, it will develop technologies to support robust inference in a multi-source knowledge transfer setting with noise-aware meta-learning and constrained inference, particularly for knowledge curated for low-resource domains or languages. The investigated technologies will be evaluated on various KG completion tasks, and several downstream tasks in areas of NLU, bioinformatics and medical informatics.

The results of the project will be disseminated by publishing papers, releasing open-source software and learning resources, organizing tutorials and workshops, and creating new courses on knowledge base construction and natural language processing.

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.