Improving accessibility and multi-hop reasoning in knowledge graphs

Hill, Blaine

Permalink

https://hdl.handle.net/2142/124280 Copy

Description

Title

Improving accessibility and multi-hop reasoning in knowledge graphs

Author(s)

Hill, Blaine

Issue Date

2024-04-18

Director of Research (if dissertation) or Advisor (if thesis)

Tong, Hanghang

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Knowledge Graph
• Knowledge Graph Reasoning
• Visualization Systems
• Transfer Learning

Abstract

Knowledge graphs have emerged as a powerful way to represent and reason over structured data and relationships. However, significant challenges persist in making the field of knowledge graph reasoning (KGR) more accessible, interpretable, and collaborative. This thesis aims to tackle these challenges through two complementary research thrusts. The first thrust introduces Ginkgo-P, a web-based platform designed to make KGR more accessible and open. Ginkgo-P provides an intuitive interface for users to visualize and interact with a wide range of core KGR algorithms spanning node recommendation, link prediction, question answering, and reinforcement learning-based reasoning. Crucially, Ginkgo-P is architected as an open platform, enabling researchers to seamlessly integrate and visualize their custom knowledge graph algorithms alongside prepackaged demonstrations. By abstracting away complexities, Ginkgo-P removes unnecessary obstacles, fostering increased collaboration and knowledge sharing within the KGR research community. The second thrust focuses on advancing the state-of-the-art in multi-hop knowledge graph reasoning using reinforcement learning techniques. A novel transfer learning approach called "Split Multi-Hop Knowledge Graph Reasoning with Reward Shaping" is proposed. This approach introduces a reward shaping mechanism that leverages pre-trained knowledge graph embeddings to estimate rewards for partial solutions, alleviating issues arising from the inherent incompleteness of knowledge graphs. Furthermore, the reward shaping framework is extended by proposing a transfer learning paradigm and investigating the integration of BERT contextualization and prompt learning techniques to improve performance by incorporating contextual information. Through the development of Ginkgo-P and the proposed Split Multi-Hop Knowledge Graph Reasoning with Reward Shaping approach, this thesis aims to contribute novel technical innovations while fostering a more open, collaborative, and accessible research environment within the KGR field.

Graduation Semester

2024-05

Type of Resource

Thesis

Copyright and License Information

Copyright 2024 Blaine Hill

Owning Collections