Research map / Large language models / Retrieval and knowledge
Learning Pattern-Based Extractors from Natural Language and Knowledge Graphs: Applying Large Language Models to Wikipedia and Linked Open Data
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
A doctoral consortium proposal plan to learn seq-to-seq (BART/REBEL-style) RDF extractors from aligned Wikipedia text and DBpedia/Wikidata graphs, with preliminary experiments on datatype properties and linearization syntaxes.
Problem
Extracting knowledge missing from DBpedia and Wikidata out of aligned Wikipedia text, producing RDF triples that follow specific SHACL-constrained patterns, especially datatype properties prone to LLM hallucination.
Approach
Fine-tune seq-to-seq transformers on a dual base of Wikipedia abstracts and their RDF graphs, comparing RDF linearization syntaxes and filtering target graphs with SHACL shapes to learn pattern-specific extractors.
Key results
- BART fine-tuned on Wikipedia abstracts and DBpedia Turtle reached F1-micro above 0.9 on dbo:birthDate, dbo:deathDate and rdfs:label for Person entities, with simplified Turtle learned fastest.
- Among seven linearization syntaxes, Turtle was easiest for BART to learn, followed by JSON-LD and RDF-XML, while N-Triples was hardest.
- A literature review suggests RE is shifting from encoder-based discriminative models toward generative encoder-decoder and decoder-only models.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use simplified Turtle instead of the other tested output syntaxes | BART · DBpedia/Wikipedia datatype-property extraction | training speed to F1-micro saturation | Helped |
| Use list or tagged-sequence output instead of the other tested syntaxes | BART · DBpedia/Wikipedia datatype-property extraction | training speed to F1-micro saturation | Helped |
| Use N-Triples output instead of the other tested syntaxes | BART · DBpedia/Wikipedia datatype-property extraction | training speed to F1-micro saturation | Hurt |
| Use Turtle output rather than JSON-LD or RDF-XML | BART · DBpedia/Wikipedia datatype-property extraction | training speed to F1-micro saturation | Helped |
Surprising findings
- N-Triples was harder for BART to learn than more verbose RDF syntaxes.
Limitations
- Preliminary experiments limited to Person entities and three DBpedia properties.
- Experiments lack cross-validation and suitable evaluation measures beyond F1-micro, as acknowledged by the author.
- Doctoral proposal; the systematic review and experiments are not yet formalized or peer reviewed.
Experimental setup
Datasets: DBpedia Wikipedia abstracts Wikidata
Metrics: F1-micro
Builds on: REBEL
Closest papers by meaning
- Well-Written Knowledge Graphs: Most Effective RDF Syntaxes for Triple Linearization in End-to-End Extraction of Relations from Texts (Student Abstract) (AAAI 2024 · similarity 0.91)
- KGGen: Extracting Knowledge Graphs from Plain Text with Language Models (NeurIPS 2025 · similarity 0.88)
- Text2NKG: Fine-Grained N-ary Relation Extraction for N-ary relational Knowledge Graph Construction (NeurIPS 2024 · similarity 0.86)
- Revisiting Document-Level Relation Extraction with Context-Guided Link Prediction (AAAI 2024 · similarity 0.85)
- Physics of Language Models: Part 3.1, Knowledge Storage and Extraction (ICML 2024 · similarity 0.85)
- An Autoregressive Text-to-Graph Framework for Joint Entity and Relation Extraction (AAAI 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.