Dependency Structure-Enhanced Graph Attention Networks for Event Detection

Qizhi Wan, Changxuan Wan, Keli Xiao, Kun Lu, Chenliang Li, Xiping Liu, Dexi LiuAAAI 2024 · DOI

TL;DR

This paper introduces a novel framework, DSE-ED, that enhances event detection by jointly leveraging token dependency trees (TDG) and a new Dependency Relation Graph (DRG). By treating dependency relations as nodes in DRG and adjusting core argument levels, DSE-ED captures richer semantic information and achieves state-of-the-art performance on event detection tasks.

What Problem Does This Paper Address?

Event detection, a crucial task in information extraction, aims to identify and classify events within text. Existing methods often struggle with three key limitations:

  1. Insufficient consideration of dependency relation structures: They primarily focus on token-level dependencies, overlooking the intricate relationships between these dependencies.
  2. Limited exploration of directed-edge semantics: The directional nature of dependency relations, which carries significant semantic meaning, is not fully exploited.
  3. Issues in strengthening core argument representation: Core arguments, vital for event identification, are often overshadowed by less important tokens due to their lower hierarchical position in dependency trees.

This paper aims to address these shortcomings by proposing a more comprehensive graph-based approach that explicitly models both token and dependency relation structures.

Methods

The proposed Dependency Structure-Enhanced Event Detection (DSE-ED) framework introduces several key innovations:

  • Dependency Relation Graph (DRG): Beyond the traditional Token Dependency Graph (TDG), DSE-ED constructs a DRG where each dependency edge in the TDG becomes a node. This allows for the aggregation and updating of dependency relation embeddings, capturing inter-relation structures.
  • Core Argument Adjustment: To boost the importance of core arguments (subjects, objects, and parallel clauses), their levels within both TDG and DRG are elevated. This helps them gain more influence during the learning process.
  • Dependency Structure-Enhanced Graph Attention Network (DSE-GAT): The framework employs Graph Attention Networks (GATs) to encode both TDG and DRG. Crucially, it refines the attention mechanism to:
  • Jointly leverage token and dependency relation embeddings: This allows for a more nuanced understanding of edge importance.
  • Incorporate directional semantics: The attention calculation considers the specific dependency relation type and the roles of neighboring tokens.
  • Enable inter-graph interaction: Node representations from TDG and DRG are updated interactively, enriching their semantic representations.

Datasets

The DSE-ED model was evaluated on the following datasets:

  • ACE2005 English corpus: A standard benchmark for event detection.
  • ACE2005 Chinese corpus: Used to demonstrate the model's cross-lingual applicability.

The StanfordCoreNLP tool was used for generating token dependency trees.

Key Insights and Ideas

  • Dependency relations are first-class citizens: Treating dependency relations as nodes in a separate graph (DRG) unlocks richer structural information that is missed by traditional TDG-only approaches.
  • Core arguments matter more than their position suggests: Elevating the levels of core arguments significantly improves event detection, highlighting their critical role and the limitations of solely relying on syntactic structure.
  • Inter-graph interaction is powerful: Allowing TDG and DRG to influence each other's representations leads to a more robust understanding of sentence semantics for event detection.
  • Directional semantics in dependency relations are crucial: Explicitly modeling the direction and type of dependency relations in the attention mechanism enhances the model's ability to discern event triggers and arguments.

Key Results

The DSE-ED model achieved state-of-the-art performance on both ACE2005 English and Chinese datasets:

  • ACE2005 English: Achieved an F1 score of 86.3%, outperforming the best benchmark by 3.5 percentage points.
  • ACE2005 Chinese: Achieved an F1 score of 79.6%, outperforming the best benchmark by 3.4 percentage points.

Ablation studies demonstrated the significant contributions of both TDG (12.2% F1 drop without it) and DRG (8.8% F1 drop without it), with TDG showing a slightly larger impact. Core argument adjustment also proved highly effective, with specific adjustments (subject, object, parallel) yielding substantial improvements.

Limitations and Future Work

  • Limitations:
    • The paper notes that DRG embeddings are initially randomly initialized and require learning, whereas BERT embeddings for tokens are pre-trained, giving TDG an initial advantage.
  • The "over-smoothing" issue, common in deep GNNs, can affect performance with too many layers.
  • Future Work:
    • The authors suggest that the hierarchical levels of core arguments are a "newly discovered clue" that could be further explored for event type identification.
  • The effectiveness of DRG in capturing better edge representations by aggregating neighbor features suggests potential for further graph-based modeling in NLP tasks.

Summary updated

BibTeX
@inproceedings{wan2024dependency,
  author = {Qizhi Wan and Changxuan Wan and Keli Xiao and Kun Lu and Chenliang Li and Xiping Liu and Dexi Liu},
  title = {{Dependency Structure-Enhanced Graph Attention Networks for Event Detection}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i17.29877},
}
Figure