Research map / Graph learning / Knowledge and temporal graphs
Coreference Graph Guidance for Mind-Map Generation
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 CMGN, a mind-map generator that builds a coreference graph over sentences, encodes it with a GCN, and uses graph contrastive learning with encoder perturbation to refine the sentence relation graph.
Problem
Automatic mind-map generation from documents requires capturing governing relations between sentences, including long-range ones; sequence-to-graph methods capture sequential features but little structure.
Approach
Introduce an explicit coreference graph (edges from the first sentence mentioning an entity to later mentioning sentences) as structural prior, encode it with a GCN, and regularize with graph contrastive learning where negatives come from a parameter-perturbed encoder rather than corrupted graphs.
Key results
- CMGN reaches 44.27 average ROUGE on SSM and 43.67 on KSM, beating EMGN (42.73 and 42.43) on the test set.
- Removing the coreference graph encoder drops SSM average to 21.74, showing the graph encoder is essential.
- CMGN outperforms ChatGPT, LLaMA2-70B and ErnieBot-4.0 on SSM/KSM (best LLM average around 45.87 R-avg? reported 52.41 vs 45.87 on a different scale), with inference of 4.09 s on the test set versus 10.27 s for EMGN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the coreference graph encoder (CGE) (ablation) | CMGN · SSM | Avg: 21.74 -> 44.27 | Helped |
| add the coreference graph encoder (CGE) (ablation) | CMGN · KSM | Avg: 24.60 -> 43.67 | Helped |
| add the graph enhancement module (GEM) (ablation) | CMGN · SSM | Avg: 42.31 -> 44.27 | Helped |
| add the graph enhancement module (GEM) (ablation) | CMGN · KSM | Avg: 42.08 -> 43.67 | Helped |
| add both CGE and GEM (ablation) | CMGN without CGE and GEM · SSM | Avg: 41.57 -> 44.27 | Helped |
| add both CGE and GEM (ablation) | CMGN without CGE and GEM · KSM | Avg: 41.55 -> 43.67 | Helped |
| replace GloVe embeddings with BERT embeddings | CMGN · SSM | Avg: 43.71 -> 39.22 | Hurt |
| replace GloVe embeddings with DistilBERT embeddings | CMGN · SSM | Avg: 43.71 -> 38.12 | Hurt |
| replace GloVe embeddings with RoBERTa embeddings | CMGN · SSM | Avg: 43.71 -> 38.27 | Hurt |
| replace GloVe embeddings with Sentence-BERT embeddings | CMGN · SSM | Avg: 43.71 -> 39.09 | Hurt |
| replace GloVe embeddings with BERT embeddings | CMGN · KSM | Avg: 43.20 -> 38.82 | Hurt |
| replace GloVe embeddings with DistilBERT embeddings | CMGN · KSM | Avg: 43.20 -> 38.92 | Hurt |
Surprising findings
- Transformer-based sentence encoders (BERT, RoBERTa, Sentence-BERT) did not improve over GloVe embeddings.
- Removing the coreference graph encoder causes a larger drop than removing the graph enhancement module.
Limitations
- LLM comparison limited to documents with at most 25 sentences due to context length.
- Coreference relations are not always relevant to the document's central subject, motivating the contrastive module.
- Relies on pseudo labels from a fine-tuned DistilBERT for training.
Experimental setup
Datasets: CNN news articles (mind-map benchmark, 135 articles) 5000 CNN articles with DistilBERT pseudo graphs
Benchmarks: Mind-map benchmark (135 CNN articles, 120 test)
Metrics: ROUGE-1 ROUGE-2 ROUGE-L average ROUGE inference time (s)
Baselines: Random LexRank MRDMF DistilBERT EMGN ChatGPT LLaMA2-70B ErnieBot-4.0
Builds on: sequence-to-graph EMGN MRDMF DistilBERT GIN
Closest papers by meaning
- Graph of Thoughts: Solving Elaborate Problems with Large Language Models (AAAI 2024 · similarity 0.84)
- A Systematic Exploration of Knowledge Graph Alignment with Large Language Models in Retrieval Augmented Generation (AAAI 2025 · similarity 0.84)
- Think-on-Graph 2.0: Deep and Faithful Large Language Model Reasoning with Knowledge-guided Retrieval Augmented Generation (ICLR 2025 · similarity 0.84)
- KGGen: Extracting Knowledge Graphs from Plain Text with Language Models (NeurIPS 2025 · similarity 0.83)
- Attention Mechanisms Perspective: Exploring LLM Processing of Graph-Structured Data (ICML 2025 · similarity 0.83)
- ICLR: In-Context Learning of Representations (ICLR 2025 · similarity 0.83)
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.