Knowledge graph reasoning and its applications: A pathway towards neural symbolic AI

Liu, Lihui

Permalink

https://hdl.handle.net/2142/125543 Copy

Description

Title

Knowledge graph reasoning and its applications: A pathway towards neural symbolic AI

Author(s)

Liu, Lihui

Issue Date

2024-07-02

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

Tong, Hanghang

Doctoral Committee Chair(s)

Tong, Hanghang

Committee Member(s)

• Zhai, ChengXiang
• Sundaram, Harri
• Sun, Yizhou

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• knowledge graph reasoning
• neural symbolic AI

Abstract

Artificial intelligence (AI) has been transforming the way we live, work, and interact with the world, and neural symbolic AI has emerged in recent years, promising next-generation AI systems that are more explainable, trustworthy, and versatile by combining the power of deep learning with symbolic reasoning, expected to revolutionize applications ranging from code generation and question answering to drug discovery; to fully unleash neural-symbolic reasoning, it is crucial to represent symbolic knowledge and integrate it with neural models, with knowledge graphs—structured representations of knowledge that capture relationships between entities and concepts in a graph-like format—providing a powerful and versatile tool for organizing and connecting real-world information; a knowledge graph (KG) is a graph-based data structure representing real-world facts in the form of triples (subject, predicate, object) and finds use in applications such as search engines, recommender systems, and question-answering (KGQA); the primary aim of knowledge graph reasoning is to extract meaningful insights from the knowledge graph data, involving discovering and explaining existing knowledge or inferring new knowledge from the graph, and reasoning on knowledge graphs with neural network models leverages the advantages of neural symbolic AI. Despite remarkable progress in knowledge graph reasoning, several challenges remain, such as dealing with the vast and incomplete nature of background knowledge graphs where explicit knowledge is stored as triplets and implicit knowledge is hidden within the graph structure, paths, or subgraphs, reasoning efficiently on large graphs, handling varying and potentially ambiguous or iterative reasoning inputs, supporting rich properties of relations in knowledge graphs including transitivity, symmetry, and asymmetry, and ensuring the model's generalization ability to solve multiple tasks and transfer to other knowledge graphs; addressing these challenges is crucial for further advancements in neural symbolic knowledge graph reasoning. In this thesis proposal, we study the problem of knowledge graph reasoning to collectively address the above challenges with neural symbolic techniques, with the completed work categorized into the following five questions. Q1: when the knowledge graph is complete and the query is accurate, how to design an efficient reasoning model? Q2: when the knowledge graph is complete and the query is uncertain, how to detect the inconsistency in the query? Q3: when the knowledge graph is incomplete and the query is accurate, how to mitigate the KG incompleteness? Q4: when the knowledge graph is incomplete and the query is ambiguous, how to find answer accurately? Q5: when the knowledge graph is incomplete and the query is dynamic, how to find answer iteratively? The first objective is to design an efficient symbolic reasoning model that aims to maximize both accuracy and speed of the query process. This involves creating a reasoning algorithm that can quickly traverse the knowledge graph, accurately identify relevant entities and relationships, and return the most precise answers. The second objective aims to detect the inconsistency of the input query with respect to the knowledge graph. This helps in verifying the truthfulness of the input query and identifying any fake information within it. The third objective aims to mitigate the incompleteness of the KG to minimize its impact on the query results. This can be achieved by leveraging the existing information in the graph, employing neural symbolic reasoning techniques to fill in missing information. The fourth objective aims to accurately find answers by leveraging limited symbolic KG information and contextual cues from the query, which involves using natural language processing techniques to understand the intent and semantics of the query, while simultaneously reasoning over the knowledge graph to get the most likely answers. The fifth objective aims to iteratively find answers by adapting the neural symbolic reasoning strategy to the changing nature of the query, and maintain both accuracy and efficiency in a constantly evolving query environment. Regarding the completed works, a symbolic subgraph matching approach has been developed to accelerate querying large knowledge graphs for Q1. For Q2, a symbolic graph kernel-based method has been proposed to detect fake information in the query. For Q3, two neural symbolic reasoning approaches have been formulated to address the graph's incompleteness. Additionally, a specific neural symbolic approach has been devised for Q4 to answer ambiguous queries, and another approach has been proposed to match ambiguous entities. Concerning Q5, a reinforcement learning-based conversational question-answering algorithm has been proposed to iteratively find answers in dynamic query environments according to the symbolic paths in the knowledge graph.

Graduation Semester

2024-08

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/125543

Copyright and License Information

Copyright 2024 Lihui Liu

Owning Collections