Research map / Graph learning / Knowledge and temporal graphs
Text2NKG: Fine-Grained N-ary Relation Extraction for N-ary relational Knowledge Graph Construction
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.
Text2NKG extracts fine-grained n-ary relational facts from text across four NKG schemas via span-tuple multi-label classification with hetero-ordered and output merging, reporting SOTA F1 on an extended HyperRED benchmark.
Problem
Automated n-ary relation extraction for n-ary relational knowledge graph construction remains coarse-grained: existing methods support only one schema, ignore entity order, and cannot handle variable arity.
Approach
Classify all ordered span-tuples of three entities with multi-label relation classifiers over packed levitated markers, augment training by permuting the 6 entity orders with relation inversion, merge the 6 orderings' probabilities at decoding (hetero-ordered merging), and then merge 3-ary facts into higher-arity facts (output merging).
Key results
- On HyperRED hyper-relational test set with BERT-base, Text2NKG reaches 83.63 F1 versus 65.04 for CubeRE (+18.60 points).
- With BERT-large, Text2NKG reaches 83.81 F1 on the hyper-relational test set versus 66.75 for CubeRE.
- Unsupervised ChatGPT and GPT-4 reach only about 11-16 F1 across the four schemas on HyperRED.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Text2NKG instead of CubeRE | HyperRED hyper-relational extraction with BERT-base · HyperRED hyper-relational / Test | F1: 65.04 ± 0.29 -> 83.63 ± 0.63 | Helped |
| Use Text2NKG instead of CubeRE | HyperRED hyper-relational extraction with BERT-large · HyperRED hyper-relational / Test | F1: 66.75 ± 0.28 -> 83.81 ± 0.54 | Helped |
| Use Text2NKG instead of LasUIE | HyperRED event-based extraction with BERT-base vs T5-base · HyperRED event-based | F1: 78.82 ± 0.26 -> 82.58 ± 0.20 | Helped |
| Use Text2NKG instead of LasUIE | HyperRED role-based extraction with BERT-base vs T5-base · HyperRED role-based | F1: 75.61 ± 0.24 -> 82.64 ± 0.38 | Helped |
| Use Text2NKG instead of LasUIE | HyperRED hypergraph-based extraction with BERT-base vs T5-base · HyperRED hypergraph-based | F1: 74.96 ± 0.42 -> 85.03 ± 0.86 | Helped |
| Add data augmentation (ablation) | Text2NKG · HyperRED hyper-relational / Test | F1: 72.06 ± 0.34 -> 83.63 ± 0.63 | Helped |
| Add null-label weight α (ablation) | Text2NKG · HyperRED hyper-relational / Test | F1: 81.97 ± 0.58 -> 83.63 ± 0.63 | Helped |
| Add hetero-ordered merging (ablation) | Text2NKG · HyperRED hyper-relational / Test | F1: 67.72 ± 0.48 -> 83.63 ± 0.63 | Helped |
Surprising findings
- Removing hetero-ordered merging drops test F1 from 83.63 to 67.72, the largest ablation effect.
- Zero-shot ChatGPT and GPT-4 perform far worse (about 11-16 F1) than supervised BERT-based models, which the authors attribute to difficulty distinguishing closely related fine-grained relations.
Limitations
- Performance is confined to supervised training; the authors note the lack of unsupervised or cold-start capability.
- Input is limited by the BERT 512-token maximum length, so long documents need a different encoder.
- The four-schema benchmark is derived from HyperRED by conversion rather than newly annotated data.
Experimental setup
Datasets: HyperRED
Benchmarks: HyperRED
Metrics: precision recall F1
Baselines: ChatGPT GPT-4 Generative Baseline Pipeline Baseline CubeRE Text2Event UIE LasUIE
Builds on: levitated markers packed levitated markers CubeRE
Closest papers by meaning
- Structure Is All You Need: Structural Representation Learning on Hyper-Relational Knowledge Graphs (ICML 2025 · similarity 0.87)
- KGGen: Extracting Knowledge Graphs from Plain Text with Language Models (NeurIPS 2025 · similarity 0.86)
- Learning Pattern-Based Extractors from Natural Language and Knowledge Graphs: Applying Large Language Models to Wikipedia and Linked Open Data (AAAI 2024 · similarity 0.86)
- Knowledge Graph Completion with Relation-Aware Anchor Enhancement (AAAI 2025 · similarity 0.86)
- HyperGraphRAG: Retrieval-Augmented Generation via Hypergraph-Structured Knowledge Representation (NeurIPS 2025 · similarity 0.86)
- 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 0.85)
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.