Research map / Graph learning / Knowledge and temporal graphs
An Autoregressive Text-to-Graph Framework for Joint Entity and Relation Extraction
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.
Proposes ATG, a span-based autoregressive text-to-graph model that generates a linearized graph of entity spans and relation triplets via a pointer over a dynamic span vocabulary, with sentence augmentation to counter premature EOS.
Problem
Joint entity and relation extraction from unstructured text, where prior generative IE models generate plain text tokens and suffer from weak grounding in the input and output structure errors.
Approach
Instead of token-level text generation, the decoder emits spans (start, end, type), relation types, and special tokens from a dynamic vocabulary built from encoder span representations, using a pointing mechanism with constrained state-transition decoding, plus sentence augmentation to reduce premature termination.
Key results
- ATG reaches 38.6 REL+ on SciERC, above the prior best and 2.1 points above the best generative approach UIE.
- ATG achieves 78.5 REL+ on CoNLL 2004, outperforming the second-best result by 2.2 points.
- On ACE 05, ATG reports 66.2 REL+ and 90.1 ENT, competitive but second-best behind ASP under its undirected evaluation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ATG instead of the strongest reported baseline | ATG · SciERC | Entity F1 | Helped |
| Use ATG instead of the strongest reported baseline | ATG · SciERC | Relation F1 | Helped |
| Use ATG instead of the strongest comparable baseline | ATG · CoNLL 2004 | Strict Relation F1 (REL+): +2.2 | Helped |
| Add positional embedding (ablation) | ATG · SciERC | Strict Relation F1 (REL+): 36.4 -> 38.6 | Helped |
| Add positional embedding (ablation) | ATG · ACE 05 | Strict Relation F1 (REL+): 66.0 -> 66.2 | Helped |
| Add positional embedding (ablation) | ATG · CoNLL 2004 | Strict Relation F1 (REL+): 78.3 -> 78.5 | Helped |
| Add structural embedding (ablation) | ATG · SciERC | Strict Relation F1 (REL+): 36.1 -> 38.6 | Helped |
| Add structural embedding (ablation) | ATG · ACE 05 | Strict Relation F1 (REL+): 65.8 -> 66.2 | Helped |
| Add structural embedding (ablation) | ATG · CoNLL 2004 | Strict Relation F1 (REL+): 78.4 -> 78.5 | Helped |
| Add positional and structural embeddings (ablation) | ATG · SciERC | Strict Relation F1 (REL+): 35.4 -> 38.6 | Helped |
| Add positional and structural embeddings (ablation) | ATG · ACE 05 | Strict Relation F1 (REL+): 65.4 -> 66.2 | Helped |
| Add positional and structural embeddings (ablation) | ATG · CoNLL 2004 | Strict Relation F1 (REL+): 78.0 -> 78.5 | Helped |
Surprising findings
- Initializing the model with a pretrained encoder-decoder performed worse than a pretrained encoder with a randomly initialized decoder, attributed to the dynamic vocabulary mismatch.
- Greedy decoding (lowest top-p) consistently performed best, despite nucleus sampling being tested.
Limitations
- Ablation of sentence augmentation and decoder depth shows sensitivity to design choices, with varying effects across datasets.
- ACE 05 results are second-best and the top ASP result is not strictly comparable due to undirected relation evaluation.
Experimental setup
Datasets: CoNLL 2004 ACE 05 SciERC
Benchmarks: CoNLL 2004 ACE 05 SciERC
Metrics: Entity F1 Relation F1 (REL) Strict relation F1 (REL+)
Baselines: DyGIE++ Tab-Seq PURE PFN UniRE TablERT HySPA TANL ASP UIE LasUIE
Builds on: TANL Pix2seq DeBERTa SciBERT
Closest papers by meaning
- Span Graph Transformer for Document-Level Named Entity Recognition (AAAI 2024 · similarity 0.87)
- Small Language Model Makes an Effective Long Text Extractor (AAAI 2025 · similarity 0.87)
- Revisiting Document-Level Relation Extraction with Context-Guided Link Prediction (AAAI 2024 · similarity 0.87)
- DocKS-RAG: Optimizing Document-Level Relation Extraction through LLM-Enhanced Hybrid Prompt Tuning (ICML 2025 · similarity 0.87)
- A Positive-Unlabeled Metric Learning Framework for Document-Level Relation Extraction with Incomplete Labeling (AAAI 2024 · similarity 0.86)
- Knowledge Graph Completion with Relation-Aware Anchor Enhancement (AAAI 2025 · similarity 0.86)
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.