Skip to main content

Luisa Werner

Wednesday 11 décembre 2024

Neural-Symbolic Integration of Knowledge Extraction and Reasoning on Graph-Structured Data

Abstract

Graph-structured data has gained significant attention in recent years due to its ability to encode relationships between entities, making it a rich data structure capable of representing complex patterns, long-chain dependencies, and cyclical structures. However, graph-structured data, such as knowledge graphs, presents major challenges that must be addressed to fully unlock its potential. Since graphs are often large, partly unstructured, and incomplete, algorithms designed for graphs need to efficiently process sparse data.

In parallel, AI research has seen a surge in the development of deep learning methods, while symbolic methods have seen fewer breakthroughs recently. Nonetheless, the explainability of symbolic methods, which are based on logic and prior knowledge, has the potential to complement the strengths of sub-symbolic methods in pattern recognition, robustness, and scalability. As a result, research on neuro-symbolic methods has gained attention, aiming to unify symbolic and sub-symbolic AI approaches.

This thesis explores how neuro-symbolic methods can be applied to graph-structured data to solve reasoning tasks, such as knowledge graph completion, more efficiently and reliably. The primary focus is on how prior knowledge, such as ontologies, can be leveraged to enhance the performance of purely sub-symbolic methods. First, this thesis investigates how prior knowledge can be integrated into a graph neural network through differentiable neural layers based on fuzzy logic. Specifically, it examines the scalability and applicability of this technique across different types of graphs. Second, a neuro-symbolic method is proposed that injects knowledge into knowledge graph embeddings by integrating a semantic reasoning engine.

Date and place

Wednesday 11 décembre at 14:00
INRIA - Montbonnot

Jury members

Thesis Advisors:

-Prof. Nabil LAYAIDA (Research Director, INRIA)

-Prof. Pierre GENEVES (Research Director, CNRS Délégation Alpes)

Reviewers:

-Prof. Fatiha SAÏS (Université Paris-Saclay)

-Prof. Farouk TOUMANI (Université Blaise Pascal - Clermont-Ferrand)

Examiners:

- Prof. Massih-Reza AMINI (Université Grenoble Alpes)

- Stefania Gabriela DUMBRAVA (ENSIIE - École Nationale Supérieure d'Informatique pour l'Industrie et l'Entreprise & Institut Polytechnique de Paris)

- Prof. Axel-Cyrile NGONGA NGOMO (Paderborn University)

Submitted on December 2, 2024

Updated on December 2, 2024