Research map / Graph learning / Knowledge and temporal graphs

Dependency Structure-Enhanced Graph Attention Networks for Event Detection

AAAI 2024 · Graph learning › Knowledge and temporal graphs · approach: Graph-based dialog systems · 4 citations (OpenAlex)

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 DSE-ED, which turns dependency edges into nodes in a dependency relation graph, promotes core-argument levels, and jointly trains token and relation GATs with a refined directed-edge attention for event detection on ACE2005.

Problem

Event detection with dependency parses ignores structure between dependency relations, underuses directed-edge semantics, and does not strengthen event core arguments.

Approach

Constructs a dependency relation graph (DRG) whose nodes are the TDG dependency edges, promotes core-argument token and relation nodes to higher levels based on dependency types, and uses an attention coefficient that combines token and relation embeddings so the two GATs interact layer by layer.

Key results

• DSE-ED achieves 86.3 F1 on ACE2005 English, versus 82.7 for the best prior baseline DualGAT (+3.5 points as reported in the abstract).
• DSE-ED achieves 79.6 F1 on ACE2005 Chinese, a 3.4-point gain over the best baseline EDM (76.2).
• Removing the TDG drops F1 to 74.1 and removing the DRG drops it to 77.5 on ACE2005 English; removing core-argument adjustment drops it to 78.2.

Measured effects reported in the paper

Surprising findings

• Removing both graphs still yields 71.2 F1, which the authors note beats some baselines, attributed to dependency relation encoding and Bi-LSTM.
• Dependency-relation attention contributes more than token attention in the interaction, with DRG layer count mattering more than TDG layer count.

Limitations

• Core argument definition relies on heuristic dependency labels (subject-predicate, predicate-object, parallel) rather than gold arguments.
• Relation node embeddings are randomly initialized, so DRG contributes less than TDG in ablations.

Experimental setup

Datasets: ACE2005 English ACE2005 Chinese

Benchmarks: ACE2005 English ACE2005 Chinese

Metrics: F1

Baselines: DualGAT SA-GRCN EE-GCN MOGANED GateGCN Seq2Seq DNR AGGED CAEE EDM

Builds on: graph attention network EE-GCN BERT Bi-LSTM

Closest papers by meaning

• Hierarchical and Incremental Structural Entropy Minimization for Unsupervised Social Event Detection (AAAI 2024 · similarity 0.86)
• Attention Mechanisms Perspective: Exploring LLM Processing of Graph-Structured Data (ICML 2025 · similarity 0.86)
• Graph Structure Inference with BAM: Neural Dependency Processing via Bilinear Attention (NeurIPS 2024 · similarity 0.85)
• A Joint Framework with Heterogeneous-Relation-Aware Graph and Multi-Channel Label Enhancing Strategy for Event Causality Extraction (AAAI 2024 · similarity 0.85)
• PANDA: Prior-guided Attentional Dual-path Architecture (NeurIPS 2026 · similarity 0.83)
• SSTAG: Structure-Aware Self-Supervised Learning Method for Text-Attributed Graphs (NeurIPS 2025 · similarity 0.83)