Ελληνικό Μεσογειακό Πανεπιστήμιο
Natural language question answering over knowledge graphs: dependency-based and hybrid approaches for Greek language applications
Abstract
dc:descriptionKnowledge graphs (KG) enable structured reasoning and highly expressive querying over complex domain information, yet they remain largely inaccessible to non-technical users because querying typically requires formal languages such as Cypher and SPARQL. This barrier is further amplified in medium-resource languages such as Greek, where rich morphology and comparatively limited Natural Language Processing (NLP) tools complicate both intent interpretation and the semantic alignment of user questions with knowledge-graph structures. This thesis addresses these limitations by proposing a unified theoretical and methodological framework for natural-language interfaces to property-graph knowledge bases, validated on Greek questions in two distinct application domains. The central objective of the framework is the reliable translation of natural-language questions into executable Cypher while preserving semantic consistency between user intent and the generated formal query. This thesis offers two core methodological contributions. First, it introduces a schema-agnostic and template-free method for translating natural-language questions into executable Cypher without labeled training data, predefined templates, or manual schema specifications. The method leverages Universal Dependencies (UD) for syntactic analysis and automatically extracts the schema of the target knowledge graph. By combining UD parses with the extracted schema, it identifies semantically plausible paths over entities, properties, and relations, and synthesizes executable graph-traversal queries accordingly. Second, the thesis develops a formal hybrid framework that integrates symbolic and neural components within a unified translation-and-execution pipeline and enforces semantic consistency between natural-language intent and the resulting graph query. The framework specifies how KG-grounded retrieval and retrieval-augmented generation can be combined while imposing hard constraints via symbolic graph querying, and it incorporates principled substitution reasoning to expand the space of feasible results under constraints. It also derives time and memory complexity bounds for the pipeline stages and empirically corroborates them through end-to-end latency measurements. These contributions of this thesis are demonstrated through two system architectures in separate domains. An Academic Information System adopts a deterministic symbolic design that translates Greek questions into Cypher through dependency parsing and schema-guided entity/relation validation, achieving high precision and recall across query types ranging from factoid retrieval to multi-hop aggregations. A Diet-Aware Recipe Recommendation System implements a neuro-symbolic architecture in which large language models are used exclusively for intent understanding and natural-language response generation, while recipe retrieval and dietary constraint enforcement are performed symbolically over a KG. The system supports multi-diet retrieval under multiple simultaneous dietary restrictions and employs dynamic ingredient substitution, yielding substantial increases in valid results. Empirical evaluation demonstrates strong performance on real-world Greek queries in the academic domain and effective handling of complex multi-constraint scenarios in the dietary domain, where ingredient substitution can double the number of valid results and user studies report high System Usability Scale scores. Overall, the thesis shows that approaches not requiring supervised training data can achieve robust performance in medium-resource languages, validates cross-domain applicability through two distinct implementations, and indicates that hybrid architectures combining symbolic precision with neural flexibility can democratize access to knowledge graphs while maintaining interpretability and evidence-based grounding.
Degree
thesis:*- Grantor dc:publisher
- Ελληνικό Μεσογειακό Πανεπιστήμιο
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Tsampos, Ioannis
- Τσάμπος, Ιωάννης
Subjects
dc:subject × 30- Τεχνητή νοημοσύνη
- Μεγάλα γλωσσικά μοντέλα (LLMs)
- Επεξεργασία φυσικής γλώσσας
- Διεπαφές φυσικής γλώσσας
- Συστήματα ερωταποκρίσεων
- Γράφοι γνώσης
- Βάσεις δεδομένων γράφων ιδιοτήτων
- Σημασιολογική ανάλυση
- Αυτόματη δημιουργία ερωτημάτων βάσεων δεδομένων
- Μετάφραση κειμένου σε Cypher
- Σημασιολογική αναζήτηση
- Παραγωγή επαυξημένης ανάκτησης (RAG)
- Artificial intelligence
- Large language models
- Natural Language Processing (NLP)
- Natural Language Interfaces
- Question answering
- Knowledge graphs
- Property graph databases
- Semantic Parsing
- Automated query generation
- Text-to-Cypher translation
- Semantic search
- Retrieval Augmented Generation
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Πληροφοριακά συστήματα
- Computer and Information Sciences
- Natural Sciences
- Information Systems
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/61256