Abstract
Graph convolutional networks (GCNs) have been successfully applied to the task of event detection. However, existing works rely heavily on a fixed syntactic parse tree structure from an external parser. In addition, the information content extracted for aggregation is determined simply by the (syntactic) edge direction or type but irrespective of what semantics the vertices have, which is somewhat rigid. With this work, we propose a novel graph convolutional method that combines an adaptive graph generation technique and a multi-channel selection strategy. The adaptive graph generation technique enables the gradients to pass through the graph sampling layer by using the ST-Gumbel-Softmax trick. The multi-channel selection strategy allows two adjacent vertices to automatically determine which information channels to get through for information extraction and aggregation. The proposed method achieves the state-of-the-art performance on ACE2005 dataset.
Introduction
Event Detection (ED) is a fundamental task in Natural Language Processing (NLP) that involves identifying event triggers and classifying them into specific event types. For instance, in the sentence "She wants to call her pregnant daughter to see if she has delivered," the word "delivered" acts as a trigger for a Life:Be-Born event.
While Graph Convolutional Networks (GCNs) have become a standard tool for capturing word relationships in ED, they typically rely on static dependency parse trees generated by external tools. This reliance introduces several limitations:
- The "Multi-hop" Gap: Relevant context words are often distant in a standard syntax tree, requiring deep GCN layers that risk "over-smoothing," where node representations become indistinguishable.
- Error Propagation: Mistakes made by the external parser are inherited by the ED model with no mechanism for correction.
- Rigid Information Flow: Standard GCNs treat edges of the same type identically, ignoring the specific semantics of the connected words.
To address these issues, we introduce AGGED (Adaptive Graph Generation for Event Detection), a framework that moves away from rigid, pre-defined syntactic trees in favor of task-specific, learned graph structures.
Methodology
The AGGED framework consists of two primary innovations: an Adaptive Graph Generation module and a Gated Multi-Channel GCN (GMC-GCN).
Adaptive Graph Generation
Instead of a fixed tree, AGGED learns to construct its own graph. It utilizes Deep Biaffine Attention to calculate edge probabilities between any two words. To ensure the discrete graph-sampling process is differentiable, we employ the ST-Gumbel-Softmax trick. This allows the model to "short-circuit" traditional syntax, creating direct links between triggers and their relevant arguments regardless of their grammatical distance.
Gated Multi-Channel GCN (GMC-GCN)
Once the graph is generated, the GMC-GCN manages information flow. Unlike standard GCNs, it features multiple information channels. A multi-channel selection mechanism uses biaffine attention to dynamically choose channels based on the semantics of the connected nodes. An information gate then modulates the volume of data passed to filter out noise.
Training Strategy
A critical component of our approach is the "Burn-in" phase. We found that initializing the graph generator randomly leads to poor performance. Instead, we pre-train the generator on a standard dependency parsing task (using the Universal Dependencies English Web Treebank) to provide a linguistically sound starting point before fine-tuning for event detection.
Experiments and Results
We evaluated AGGED on the ACE2005 dataset, the standard benchmark for event detection, containing 33 predefined event types across 599 documents.
Performance
AGGED achieved state-of-the-art results, significantly outperforming previous leading models:
- F1 Score: 79.9% (compared to 77.6% for EE-GCN and GatedGCN).
- Recall: 82.2%, indicating that the adaptive graph is superior at capturing relevant context than static trees.
Ablation Study
Ablation tests confirmed the necessity of each component. Removing the adaptive graph generation (–AG) resulted in a 1.6% drop in F1 score, while removing the information gate (–IG) led to a 1.2% decrease. These results demonstrate that both structural flexibility and controlled information flow are vital for high-performance event detection.
Conclusion
AGGED demonstrates that the optimal graph for event detection is not necessarily a perfect grammatical tree. By allowing the model to learn task-specific topologies and semantic-aware information channels, we can overcome the limitations of static dependency parsing. Future work will explore relaxing the 1-indegree constraint for generated graphs and applying this architecture to broader information extraction tasks like entity and relation recognition.
Summary updated