📄 Ontology-Guided Synthetic Data Generation for Low-Resource Information Extraction: A Case Study in IT Heritage Domain

Summary

Building NER and RE systems for niche fields is hard because expert-labelled data is scarce. Existing Reverse-IE methods create synthetic training text from knowledge graphs, but they assume such a knowledge base already exists. The authors drop that assumption and start from a formal ontology instead. Their pipeline samples small graph "motifs" step by step. It picks an anchor type, draws node out-degrees from a Poisson distribution, chooses relations the schema permits, and sometimes reuses existing nodes so that paths and cycles appear. Real entity names then replace the abstract nodes, and an LLM turns each motif into text through the KGAST framework. Span-level labels follow automatically from this process. Because sampling is explicit, rare types or relations can be deliberately boosted. The method was tested on the IT Heritage domain against a small pseudo-gold set. The synthetic corpus turned out repetitive and structurally simpler than real documentation, but its volume makes it a workable way to bootstrap extractors where gold data is lacking. This matters for any domain that has a schema but no annotated corpus.

Key points

• The source schema is the IT in Heritage ontology, which has 329 classes and 86 object properties. It was trimmed to 18 core entity types and 15 relations by collapsing specialised subclasses into their parents.
• The authors produced three sets of 3,000 motifs, varying the node re-use probability α (0.3, 0.5, 0.7) with target size N = 8 and λ = 2. The α = 0.7 set was chosen for verbalization with Mistral-Small-3.1-24B-Instruct.
• Verbalization fidelity was high: the generated texts kept 94.63% of input entities and 93.45% of input triples, so roughly 6.5% of labels may be missing from the text.
• The evaluation baseline is a pseudo-gold set of 100 documents selected by LLM-as-judge reranking and labelled by an automated extractor. Checked on 10 manually annotated samples, the extractor scored precision 0.745, recall 0.991 and micro-F1 0.85. A bert-base-uncased model was fine-tuned for NER and RE under four regimes: 80 gold docs, full synthetic, random 50% and stratified 50%. All were tested on 20 held-out docs.
• The main weaknesses are high Self-BLEU (repetitive text), a small entity pool and sparse motif topology, which leads to a "one fact per sentence" style. Planned fixes are a real expert gold benchmark, denser motif objectives and a larger, distribution-aware entity pool.