Kategorie: ‘Theses’
A Generative Foundation Model for Knowledge Graphs: Geometry- and Text-Aware Pretraining for Transferable Downstream Tasks
Knowledge graphs like Wikidata combine rich relational structure with natural-language descriptions, yet most models are trained narrowly for a single task and transfer poorly. This thesis investigates how a single generative graph foundation model, pretrained on large-scale text-rich knowledge graphs, can be adapted to a range of downstream tasks, including knowledge graph completion, text-conditional subgraph generation, and graph anomaly detection, with minimal task-specific supervision. The work integrates geometry-aware representation learning, text-conditioned graph transformers, and generative graph modelling into one transferable pretraining-and-adaptation pipeline.
Ontology-Based Data Augmentation with LLMs for Narrative Classification
Narrative Classification identifies stories via NLP but often lacks generalizability. While LLMs augment other text tasks, their narrative application remains exploratory. This thesis investigates whether an ontology-based LLM-agent framework incorporating specific data characteristics improves synthetic training data quality.
Traceability Framework for Human–LLM-Assisted Tabular Data Transformations
Large Language Models (LLMs) are increasingly used to support data wrangling, but their integration into interactive transformation workflows raises new challenges for auditability, reproducibility, and accountability. When users approve, reject, or refine LLM-generated suggestions, conventional data lineage systems often fail to capture why a change occurred, who was responsible for it, and which transformation produced the final dataset.
This thesis investigates a compact traceability framework for human–LLM-assisted transformations of uploaded tabular files. The target setting is a single structured tabular data file (e.g. CSV), column-level transformation workflows, and practical reproducibility. The framework tracks file versions, table versions, selected columns, LLM suggestions, human decisions, approved transformation specifications, generated code references, execution events, and resulting output versions, with the goal of enabling reconstruction and rollback without storing the full conversation.
This paper is co-supervised by Prof. Jiannan Wang (jnwang@tsinghua.edu.cn) from Department of Computer Science and Technology at Tsinghua University , who also serves as the second supervisor.
Training a Tiny LLM with Block Attention Residuals on CommonsenseQA
Knowledge-augmented multiple-choice question answering (MCQA) aims to improve robustness and factual grounding by integrating external structured knowledge (e.g., knowledge graphs) into language-model-based decision making. Current high-performing systems typically retrieve a local subgraph relevant to a question and candidate answers, then combine pretrained language representations with explicit graph reasoning modules.
This thesis investigates an alternative representation path: instead of processing retrieved knowledge graph (KG) subgraphs as symbolic triples with graph neural networks, the subgraphs are deterministically rendered into a compact 2D “visual graph” representation and encoded with a vision backbone. The resulting visual KG evidence is fused with a encoder-only language model via attention-based cross-modal interaction. The core research question is whether a visually encoded KG can preserve decision-relevant relational structure and support competitive knowledge-augmented MCQA performance on CommonsenseQA and OpenBookQA (optionally extending to MedQA-USMLE).
Master’s Thesis: Automatic Generation of Ontological Representations for Co-Simulation Configurations in Power Engineering
Abstract
This master’s thesis aims at examining the applicability of automatic ontology generation and ontology-based data integration to the configuration of co-simulation scenarios. To study power systems through simulations, it is conducive to model sub-domains through separate simulators, which are combined through co-simulations to comprise complex simulation scenarios. However, what is gained through focused modelling of subdomains is paid in complexity, when configuring scenarios comprising of many and heterogenous simulators. Some use cases also necessitate the integration of co-simulations with external systems such as data bases or user-interfaces, which needs to be reflected in the configuration of the simulations. Mechanisms for validation and rule-based configuration are necessary. The student will contribute to a framework for ontology-based configuration of co-simulation scenarios by exploring the possibility of automatically generating ontologies for simulators and integrating these with another and with external ontologies for integration with external systems.
Ontology-Grounded Extraction of Research SoftwareMentions from Scientific Publications
Research software is among the least discoverable scholarly outputs. While standards like CodeMeta and CFF enable structured software metadata at the repository level, they require active curation by maintainers and see inconsistent adoption. On the publication side, only select publishers such as Schloss Dagstuhl’s DROPS platform provide citable software artifacts, again contingent on explicit author action. As a result, most research software is mentioned only in unstructured publication text, invisible to metadata-driven search and agentic systems, and non-compliant with FAIR principles. This problem is compounded in applied domains like visualization research, where publications frequently describe custom tools and prototypes that are never publicly released, a case not covered by current metadata schemas. Automating ontology-grounded extraction of software mentions from publications would close this gap and enrich the metadata foundation for downstream services such as the research copilots developed within NFDI4DS.
Incremental Knowledge Graph Ingestion with Change Detection and Provenance Tracking
Keeping a knowledge graph up to date as its source data evolves is harder than building one from scratch. New records appear, existing records are corrected, and metadata is enriched over time. Each type of change a corrected DOI, an added co-author, a retracted publication carries different semantic implications and may require a different update strategy. Detecting these changes efficiently and propagating them without introducing duplicates, losing provenance, or overwriting valid data remains an open challenge, particularly when the goal is to avoid heavyweight versioning infrastructure.
From Triples to Pixels: Visual Knowledge Graph Encoding for Knowledge-Augmented Multiple-Choice Question Answering
Knowledge-augmented multiple-choice question answering (MCQA) aims to improve robustness and factual grounding by integrating external structured knowledge (e.g., knowledge graphs) into language-model-based decision making. Current high-performing systems typically retrieve a local subgraph relevant to a question and candidate answers, then combine pretrained language representations with explicit graph reasoning modules.
This thesis investigates an alternative representation path: instead of processing retrieved knowledge graph (KG) subgraphs as symbolic triples with graph neural networks, the subgraphs are deterministically rendered into a compact 2D “visual graph” representation and encoded with a vision backbone. The resulting visual KG evidence is fused with a encoder-only language model via attention-based cross-modal interaction. The core research question is whether a visually encoded KG can preserve decision-relevant relational structure and support competitive knowledge-augmented MCQA performance on CommonsenseQA and OpenBookQA (optionally extending to MedQA-USMLE).
A Rule-Based Agent for Semantic Matching Graph Visualization
Resource-Efficient Cyber Risk and Criticality Assessment for Small Power Grid Operators: A Reproducible Algorithm for Deriving Security Requirements and Prioritized Mitigation Plans
Power grids are increasingly operated through tightly interconnected IT/OT infrastructures, which raises the attack surface and makes smaller operators with limited resources particularly vulnerable to security-relevant incidents. This thesis develops and evaluates a reproducible, resource-efficient analysis algorithm that captures essential system, process, role, location, and information-flow data to derive protection needs and criticality, and to generate a prioritized, actionable security improvement plan without requiring a full ISMS. The approach is prototyped and validated through a realistic case study, benchmarking against at least one reference method and incorporating expert interviews to assess effort, comprehensibility, traceability, domain coverage, and prioritization quality.
Master Thesis