Dissertation
Developing novel triple embeddings for scalable alignment of knowledge graphs and natural language
Doctor of Philosophy (Ph.D.), Drexel University
Jan 2024
DOI:
https://doi.org/10.17918/00001906
Abstract
Neural representation learning has become the de facto mode of large-scale information compression over the past decade. These compression schemes generate embeddings: dense, floating-point vector representations that aim to capture semantic relatedness between data elements. These approaches have been applied in both the natural language and knowledge graph representation settings, but with the rise and pervasive nature of large language models, a critical eye has been turned to the black box nature of these algorithms. Proposals have been generated for the use of a hybrid scheme, where language models and knowledge graphs operate in a unified way, yet little work exists in how to tie the internal representational states of these approaches together. This problem can be viewed as a distant supervision problem, where entities and relations are extracted from natural language into knowledge graph triple format. This dissertation attempts to bridge the unification gap via distant supervision, first by experimenting with linear methods to learn mappings between representations in each respective domain. Through these experiments, we discovered two main issues: the lack of predicate, or relational, information in knowledge graph representations, and the robustness to noise and scalability concerns in the mapping procedure. To address these concerns, we introduce a novel triple representation model based on pairwise triple similarity scoring (PTSS). Using these representations, we then introduce a batched information guided optimal transport (BIG-OT) to align natural language and knowledge graph representations, achieving state-of-the-art results in both accuracy and run-time on two benchmark distant supervision tasks.
Metrics
99 File views/ downloads
75 Record Views
Details
- Title
- Developing novel triple embeddings for scalable alignment of knowledge graphs and natural language
- Creators
- Alexander Kalinowski
- Contributors
- Yuan An (Advisor)
- Awarding Institution
- Drexel University
- Degree Awarded
- Doctor of Philosophy (Ph.D.)
- Publisher
- Drexel University; Philadelphia, Pennsylvania
- Number of pages
- xii, 153 pages
- Resource Type
- Dissertation
- Language
- English
- Academic Unit
- Computer Science (Computing) (2013-2026); College of Computing and Informatics (2013-2026); Drexel University
- Other Identifier
- 991021823312804721