Background
Within the JudaicaLink project, our group builds and maintains a knowledge graph for the domain of Jewish Studies, integrating and interlinking reference works, authority data, and digitized collections. One of the core resources is Compact Memory, a large newspaper archive of more than one million digitized pages, mostly from the 19th century. The pages exist as scanned images, and OCR text was produced and processed for named entity recognition and full-text search as part of an earlier project phase — but this processing dates back to around 2020, before the current generation of large language models (LLMs) was available. The OCR quality is often poor, as is typical for historical newspapers with period typefaces, degraded originals, and complex layouts.
This creates an interesting opportunity: modern LLMs and vision-language models (VLMs) have substantially changed what is feasible for both text recognition and information extraction on noisy, historical material. This thesis should explore how much can be gained by revisiting this corpus (or a representative subset of it) with current technology, and how the results can be meaningfully integrated into the JudaicaLink knowledge graph.
Research Questions
The thesis should address questions on two levels:
Text and image processing:
- How much additional, usable information can be extracted from the existing (2020-era) OCR text using current LLMs, compared to what was extracted at the time?
- Would a fresh OCR pass be worthwhile? Our group has had good experience using LLMs to work with noisy text of the kind produced by OCR — does this also hold for post-correcting OCR output at scale, or for re-recognizing the text from scratch?
- Alternatively, is it preferable to skip a dedicated OCR step altogether and process the page images directly with a vision-language model, potentially combining recognition and extraction in one step?
Knowledge graph construction:
- How can entities found in the text be reliably matched against entities already present in the JudaicaLink knowledge graph (and relevant authority files, e.g. the GND)?
- When should a new entity be created instead, and how should such decisions be made and validated?
- How can existing entities be enriched with information gathered from the corpus?
- How should the provenance of extracted facts be represented — i.e., how do we record where (which page, which article, which passage) a piece of information about an entity was found, so that it remains traceable and citable?
Proposed Methodology
The following is conventiently created for you by AI so you don’t have to do it. Check it, improve it, ignore it, whatever is appropriate.
The following is one possible way to structure the work — how exactly to approach the topic, and which parts to prioritize, is deliberately left open and should be worked out by the student based on their own literature review during the exposé phase. A Master’s thesis will typically not be able to cover all aspects in equal depth, so an early scoping decision (e.g. focusing more on the text/image processing side, or more on the knowledge graph side) is expected and encouraged.
A rough sketch:
- Baseline and corpus selection. Define a manageable, representative sample of the corpus (or a comparable historical text collection) together with some form of gold standard or manual spot-checking, to allow meaningful comparison between approaches.
- Text quality / extraction pipeline. Compare at least two of: (a) LLM-based post-correction of the existing OCR output, (b) re-OCR with a current OCR engine, possibly followed by LLM correction, (c) direct extraction from page images using a vision-language model. Evaluate not only text quality (e.g. character/word error rate) but downstream extraction quality.
- Information extraction. Extract entities (and possibly relations) relevant to the JudaicaLink domain (persons, places, organizations, publications, dates, subjects, etc.), adapted to the existing JudaicaLink ontology.
- Knowledge graph integration. Design and evaluate a strategy for matching extracted entities against existing knowledge graph entities and authority data, for creating new entities where appropriate, for enriching existing entities, and for representing the provenance/evidence (“Fundstellen”) of extracted statements (e.g. via reification, named graphs, or a provenance vocabulary such as PROV-O).
Getting Started: Selected Literature
The following references are meant as entry points into the relevant literature, not as a complete or binding reading list. Part of the thesis work is to research the current state of the art independently and to identify the most relevant and recent work yourselves.
OCR, post-correction, and document image processing with LLMs/VLMs
- Boros, E., Ehrmann, M., Romanello, M., Najem-Meyer, S., & Kaplan, F. (2024). Post-correction of historical text transcripts with large language models: An exploratory study. LaTeCH-CLfL 2024. — Reports mostly negative results for prompt-based LLM OCR correction on historical text; a useful counterpoint.
- Thomas, A., Gaizauskas, R., & Lu, H. (2024). Leveraging LLMs for post-OCR correction of historical newspapers. LT4HALA @ LREC-COLING 2024. — Fine-tunes an LLM specifically for OCR correction of historical newspapers.
- Ghiriti, A., Göderle, W., & Kern, R. (2024). Exploring the Capabilities of GPT4-Vision as OCR Engine. In: Linking Theory and Practice of Digital Libraries. — Direct image-to-text extraction with a multimodal model.
- Kim, S., Baudru, J., Ryckbosch, W., Bersini, H., & Ginis, V. (2025). Early evidence of how LLMs outperform traditional systems on OCR/HTR tasks for historical records. arXiv:2501.11623.
- Multimodal LLMs for OCR, OCR Post-Correction, and Named Entity Recognition in Historical Documents (2025). arXiv:2504.00414. — Covers all three steps in one pipeline, close to the setting of this thesis.
- OCR Error Post-Correction with LLMs in Historical Documents: No Free Lunches (2025). arXiv:2502.01205. — Good overview of the (mixed) evidence for and against LLM-based post-correction.
Named entity recognition and information extraction from historical text
- Ehrmann, M., Romanello, M., Najem-Meyer, S., Doucet, A., Clematide, S., Faggioli, G., Ferro, N., Hanbury, A., & Potthast, M. (2022). Extended overview of HIPE-2022: Named entity recognition and linking in multilingual historical documents. CLEF 2022. — Central shared-task/benchmark series for NER and entity linking on historical newspapers, a good starting point for evaluation design.
- Named Entity Recognition of Historical Text via Large Language Model (2025). arXiv:2508.18090.
Entity matching, entity linking, and knowledge graph construction
- Linhares Pontes, E., et al. (2022). MELHISSA: A multilingual entity linking architecture for historical press articles. International Journal on Digital Libraries.
- Scharpf, P., Breitinger, C., Spitz, A., Meuschke, N., Greiner-Petter, A., Schubotz, M., & Gipp, B. (2026). Entity Linking with Wikidata: A Systematic Literature Review. ACM Computing Surveys. — Broad, recent overview of entity linking approaches against a large knowledge graph, transferable to matching against JudaicaLink/GND.
- Peeters, R., Steiner, A., & Bizer, C. (2024). Entity Matching using Large Language Models. arXiv:2310.11244. — From the entity matching group at the University of Mannheim; directly relevant to the “matching found entities with existing entities” question.
- Wang, T., et al. (2024). Match, Compare, or Select? An investigation of large language models for entity matching. arXiv:2405.16884.
- LLM-empowered knowledge graph construction: A survey (2025). arXiv:2510.20345. — Broad recent survey covering LLM-based extraction, schema/ontology construction, and canonicalization.
What We’re Looking For
- Interest in NLP / information extraction and/or knowledge graphs; willingness to work with historical and imperfect data.
- Basic experience with Python and, ideally, some exposure to LLM APIs or vision-language models, as well as RDF/SPARQL.
- No prior domain knowledge is required — for specific questions, access to domain experts will be available.
Interested? Please get in touch — see the process description on our thesis page for how to proceed (initial contact, exposé, etc.).