Research map / Graph learning / Knowledge and temporal graphs
Well-Written Knowledge Graphs: Most Effective RDF Syntaxes for Triple Linearization in End-to-End Extraction of Relations from Texts (Student Abstract)
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.
Compares seven output syntaxes (list, tagged sequence, Turtle, RDF-XML, N-Triples, JSON-LD, simplified Turtle) for linearizing triples when fine-tuning BART-base and BART-large on datatype-property extraction from Wikipedia abstracts.
Problem
Output syntax choice for generating triples in end-to-end relation extraction with seq2seq models is rarely studied, and few methods produce RDF-standard knowledge graphs.
Approach
Treat the linearization of a triple set as a design choice and systematically compare RDF syntaxes (Turtle, RDF-XML, N-Triples, JSON-LD) and a simplified Turtle variant against existing list and tagged formats, across BART model sizes.
Key results
- Simplified Turtle outperformed all other syntaxes in every aspect (fast to learn, short token length of avg 58 tokens).
- N-Triples was the clear loser, taking two to three times more training epochs to reach micro-F1 saturation than other syntaxes.
- The performance gap between syntaxes narrowed with BART-large compared to BART-base; original Turtle needed the same number of epochs for both models.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use simplified Turtle syntax instead of the other evaluated syntaxes | BART-base and BART-large · DBpedia datatype relation extraction from Wikipedia abstracts | overall performance | Helped |
| Use N-Triples instead of the other evaluated syntaxes | BART-base and BART-large · DBpedia datatype relation extraction from Wikipedia abstracts | training epochs to micro-F1 saturation: two or three times longer | Hurt |
| Increase model size from BART-base to BART-large | BART · DBpedia datatype relation extraction from Wikipedia abstracts | micro-F1 performance gap between syntaxes | Helped |
| Increase model size from BART-base to BART-large | BART · DBpedia datatype relation extraction from Wikipedia abstracts | training epochs to reach micro-F1 greater than 0.9: greater than 0.9 | Helped |
| Use BART-large instead of BART-base | Turtle syntax · DBpedia datatype relation extraction from Wikipedia abstracts | training epochs | No clear effect |
| Use BART-large instead of BART-base | List syntax · DBpedia datatype relation extraction from Wikipedia abstracts | training epochs to learn the syntax | Hurt |
| Use BART-large instead of BART-base | JSON-LD syntax · DBpedia datatype relation extraction from Wikipedia abstracts | training epochs to learn the syntax | Helped |
| Use BART-large instead of BART-base | XML syntax · DBpedia datatype relation extraction from Wikipedia abstracts | training epochs to learn the syntax | Helped |
Surprising findings
- BART-large needed fewer epochs to reach good results despite taking longer per epoch, and the gap between syntaxes narrowed with the larger model.
Limitations
- Only datatype properties on dbo:Person instances were studied; object properties not evaluated.
- Evaluation uses a single split of random subsets rather than k-fold cross-validation, and the metrics may not be adequate for generated triples.
- Only two BART sizes were tested.
Experimental setup
Datasets: DBpedia English chapter (dbo:Person datatype triples) Wikipedia abstracts
Benchmarks: custom DBpedia/Wikipedia abstract datatype-triple dataset
Metrics: micro-F1 epochs to reach micro-F1 > 0.9 token length
Baselines: list syntax tagged sequence syntax
Builds on: REBEL
Closest papers by meaning
- Learning Pattern-Based Extractors from Natural Language and Knowledge Graphs: Applying Large Language Models to Wikipedia and Linked Open Data (AAAI 2024 · similarity 0.91)
- KGGen: Extracting Knowledge Graphs from Plain Text with Language Models (NeurIPS 2025 · similarity 0.87)
- Knowledge Graph Completion with Relation-Aware Anchor Enhancement (AAAI 2025 · similarity 0.86)
- Reliable Data Generation and Selection for Low-Resource Relation Extraction (AAAI 2024 · similarity 0.86)
- An Autoregressive Text-to-Graph Framework for Joint Entity and Relation Extraction (AAAI 2024 · similarity 0.86)
- Text2NKG: Fine-Grained N-ary Relation Extraction for N-ary relational Knowledge Graph Construction (NeurIPS 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.