Research paper · TU Dresden · 2026 - present
An experimental framework that turns long-form text into structured ontologies, merges them hierarchically, and benchmarks their semantic and structural quality.
As part of an ongoing research paper at TU Dresden, I am developing a framework for extracting ontologies from long documents with large language models. The system splits source material into manageable sections, generates sub-ontologies in several formats, combines them with configurable merge strategies, and evaluates the result through deterministic checks and LLM-assisted review.
Industrial knowledge is often scattered across manuals, process descriptions, notes, and other documents. Modeling that information manually as an ontology is expensive, while processing an entire long document in one LLM prompt tends to lose detail and exceeds the practical context limits of smaller models.
Built a sentence-preserving chunking and generation pipeline that creates focused sub-ontologies in OWL, Turtle, JSON, or YAML. The fragments can then be combined through top-down, bottom-up, tree-based, or sequential merging, with optional overview ontologies and an improvement loop for repairing structural and semantic issues.
The project provides a reproducible environment for comparing language models, serialization formats, chunk sizes, and merge strategies. The accompanying paper investigates how those choices affect syntactic validity, retained detail, semantic quality, and the usefulness of generated ontologies for industrial knowledge management.
Python modules handle sentence-aware chunking, parallel LLM generation, format conversion, hierarchical merging, automated improvement, and benchmark orchestration. A Next.js dashboard provides visual configuration, live process output over Server-Sent Events, and an explorer for generated ontologies and evaluation reports.
Generates and converts ontologies across OWL/XML, Turtle, JSON, and YAML to compare model behavior and representation overhead.
Supports top-down, bottom-up, tree-based, and sequential strategies for reconstructing one ontology from document-level fragments.
Combines deterministic graph checks with LLM-assisted semantic review, competency questions, and hallucination analysis.
A Next.js interface for configuring experiments, streaming live runs, and inspecting generated ontologies and benchmark reports.
Implemented document chunking, multi-format ontology extraction, and provider-independent LLM access.
Added configurable hierarchical and sequential methods for combining generated sub-ontologies.
Built structural, semantic, competency-question, and gold-standard benchmarks across models and configurations.
Running comparative experiments and developing the findings into a research paper on ontology learning for manufacturing and logistics.
