National Technical University of Athens (NTUA)
Knowledge graphs and large language models for explainable, reliable, and efficient complex data management and retrieval
Abstract
dc:descriptionModern data ecosystems are dominated by complex, large-scale, and highly interconnected information, yet most organisations still manage and retrieve this data through siloed systems or structurally opaque Artificial Intelligence (AI) models. Structured technologies such as databases and standalone knowledge graphs (KGs) provide reliability and precision, but struggle to cope with multimodal, rapidly evolving data and with natural-language access. Conversely, Large Language Models (LLMs) offer flexible semantic reasoning and fluent interaction, but their internalized knowledge is opaque, hard to update, and prone to hallucinations, making them difficult to trust in high-stakes settings. This thesis addresses the resulting core problem: how to design explainable, reliable, and efficient methods for complex data management and retrieval by tightly integrating knowledge graphs and LLMs. Using sustainable development and climate-change data as a demanding application domain, the work investigates how graph-based modelling and multimodal integration, combined with KG-grounded, agentic LLMs, can turn fragmented data into grounded, traceable, and reusable knowledge. To tackle this problem, the thesis first develops a principled Labeled Property Graph (LPG)-based KG engineering methodology and instantiates it in SustainGraph, a large-scale KG for sustainable development and climate change. SustainGraph models the multi-level governance and data landscape around the United Nations (UN) Sustainable Development Goals (SDGs), the European Green Deal, national commitments, and contextual socio-economic and environmental indicators. A modular multimodal integration pipeline harmonises temporal, spatial, and textual data from heterogeneous providers (e.g., UN SDG database, Eurostat, Copernicus, World Bank) into a unified graph representation, covering time-series indicators, geospatial data, and policy documents. This pipeline includes schema design, semantic alignment, geospatial upscaling strategies, and automated, workflow-based graph population, yielding a Findable, Accessible, Interoperable, Reusable (FAIR)-compliant, incrementally maintainable KG that serves as a reliable backbone for downstream analytics and LLM-based retrieval.Building on this infrastructure, the thesis introduces two large-scale analytical frameworks that demonstrate how KG-based modelling can support robust, reproducible reasoning over complex sustainability data. The first framework analyses interlinkages, synergies, and trade-offs among SDG indicators across national and regional spatial resolutions by automatically selecting sufficiently supported time windows, performing correlation-based network analysis, and reintegrating the discovered SDG goal–goal and indicator-indicator relations back into SustainGraph. The second proposes an open, interoperable climate change vulnerability assessment framework that leverages the KG to curate indicators across social, economic, environmental, and political dimensions, perform systematic data quality assessment and imputation, and compute multi-level, hazard-independent vulnerability indices from country to city scale. Both frameworks are implemented and evaluated on real European data, resulting in interpretable maps, clusters, and sensitivity analyses that reveal critical patterns in SDG interdependencies and regional vulnerabilities. To address the limitations of natural language access to these complex structures, the thesis introduces three tightly coupled contributions. First, CyVer, a deterministic, multi-level Cypher query validation library, decomposes query assessment into syntactic, structural, and property-level dimensions, producing structured, interpretable error metadata that precisely identifies not only whether a generated query is incorrect, but where and why it fails — a diagnostic capability that standard text similarity metrics systematically cannot provide. Second, building on CyVer, the thesis introduces a Metric-Based LLM Evaluation Framework that conducts a comprehensive benchmarking study across thirteen models of varying scale and architecture, combining CyVer's validation scores with execution-based and similarity metrics into a unified, multi-dimensional assessment protocol. This framework establishes that model scale alone is not a reliable predictor of Text-to-Cypher reliability, that failure modes differ qualitatively between smaller and larger models, and that schema representation format is a significant and underappreciated determinant of query generation quality. Third, CyVerACT, an agentic Text-to-Cypher workflow, closes the loop by moving beyond single-shot generation toward an iterative, self-correcting paradigm. By exploiting CyVer's structured error metadata as a deterministic feedback signal within a closed refinement loop augmented by dynamic schema filtering, CyVerACT enables models to progressively resolve query errors across successive iterations without reliance on human annotation or model fine-tuning. Crucially, grounding the feedback loop in deterministic, rule-based validation rather than stochastic LLM-based assessment proves essential: CyVerACT consistently outperforms both single-shot generation and iterative agentic workflows that rely on LLM-as-a-judge evaluation across all evaluated models and benchmark datasets. These results demonstrate that agentic refinement grounded in deterministic validation is a scalable and accessible pathway to high-fidelity KG query generation — one that allows open-source models to match or exceed the performance of frontier proprietary systems. Taken together, the contributions of this thesis establish a coherent, end-to-end architecture for knowledge-grounded data management: from principled KG engineering and multimodal integration, through reproducible large-scale analytics, to interpretable and auditable natural language querying. Across all layers, explainability is embedded by design — construction pipelines are traceable and schema-validated, analytical results are reproducible and sensitivity-tested, and query generation is governed by deterministic, inspectable correction logic rather than opaque model internals. This thesis thus demonstrates that the tension between the flexibility of LLMs and the rigour of structured knowledge representations is not irresolvable: through tight, principled integration, the two paradigms are mutually reinforcing, and together constitute a reliable, explainable foundation for evidence-based decision-making in complex, data-intensive domains.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Androna, Christina-Maria
- Ανδρωνά, Χριστίνα-Μαρία
Subjects
dc:subject × 18- Γράφος Γνώσης
- Μεγάλα γλωσσικά μοντέλα (LLMs)
- Τεχνητή νοημοσύνη
- Σημασιολογική μοντελοποίηση
- Συστήματα ερωταποκρίσεων
- Ερμηνευσιμότητα
- Knowledge graphs
- Large language model (LLMs)
- Artificial intelligence
- Semantic modelling
- Knowledge graph question answering (KGQA)
- Explainability
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Μηχανική και συστήματα επικοινωνιών, Τηλεπικοινωνίες
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Communication engineering and systems, Telecommunications
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/62234