Abstract
Event Detection (ED) is a critical task in information extraction that involves identifying event triggers and classifying them into predefined types. Traditional sequence-based models often struggle with long-range dependencies and irrelevant context noise. This paper proposes a novel architecture, GCN-ED, which leverages Graph Convolutional Networks (GCNs) to perform convolutions over syntactic dependency trees. By incorporating a novel "Argument-Aware" pooling strategy that focuses on entity mentions, the model effectively captures informative linguistic structures while filtering out noise. Experimental results on the ACE 2005 and TAC KBP 2015 benchmarks demonstrate that GCN-ED achieves state-of-the-art performance, highlighting the synergy between sequential and structural modeling.
Introduction
In Natural Language Processing (NLP), Event Detection (ED) is the task of identifying "event triggers"—typically verbs or nouns that most clearly express an event's occurrence—and categorizing them into specific subtypes. For instance, in the sentence "The police officer who fired into the car was fired Tuesday," a robust ED system must distinguish between the first "fired" (an Attack event) and the second "fired" (an End-Position event).
The primary challenge in ED is context. While traditional Convolutional Neural Networks (CNNs) are effective at capturing local patterns through $k$-grams, they often fail when the crucial disambiguating context is distant from the trigger. Conversely, while Recurrent Neural Networks (RNNs) like LSTMs can model longer sequences, they may incorporate irrelevant "noise" from the entire sentence. The authors argue that the grammatical structure, or syntax, of a sentence provides the necessary direct links between triggers and their informative context, regardless of linear distance.
Methodology
The proposed GCN-ED model utilizes a multi-layered architecture designed to integrate both sequential and structural information:
- Encoding and BiLSTM Layer: Input words are represented as embeddings combining word, position, and entity type information. A Bidirectional LSTM (BiLSTM) processes these embeddings to capture global sequential context and the "flow" of the sentence.
- Graph Convolutional Networks (GCNs): The core of the model is a GCN layer that operates on the sentence's dependency parse tree. Unlike standard convolutions that slide over a linear window, GCNs perform convolutions over the graph defined by grammatical relationships (e.g., subject-verb links). This allows the model to aggregate information from syntactically related words in just a few hops, effectively bypassing irrelevant intermediate clauses.
- Edge Weighting: To account for potential errors in automated dependency parsers, the model learns to assign weights to edges, prioritizing more reliable syntactic links.
- Argument-Aware (ENTITY) Pooling: Instead of standard max-pooling over the entire sentence, the authors introduce a targeted pooling strategy. The model pools representations specifically from the trigger word and the entity mentions (e.g., persons, vehicles, locations) present in the sentence. Since entities are the likely participants (arguments) of an event, this strategy acts as a natural attention mechanism, focusing the model on the most informative signals.
Experiments and Results
The model was evaluated on two standard benchmarks: ACE 2005 (using gold-standard entity annotations) and TAC KBP 2015 (using predicted, potentially noisy entity mentions).
On the ACE 2005 dataset, GCN-ED achieved an F1 score of 73.1%, outperforming previous state-of-the-art models, including non-consecutive CNNs (NCNN) and attention-based models (ATT+).
On the more challenging TAC KBP 2015 dataset, the model achieved an F1 of 58.8%, surpassing the top-ranked system from the official competition. Ablation studies confirmed that both the GCN layers and the ENTITY pooling strategy are vital for performance; removing either component led to a significant drop in F1 scores.
Conclusion
GCN-ED represents the first successful application of Graph Convolutional Networks to Event Detection. By combining the sequential modeling strengths of BiLSTMs with the structural insights of GCNs over dependency trees, the model effectively captures long-range linguistic dependencies. The introduction of Argument-Aware pooling further refines the model's focus on relevant participants, setting a new standard for the task. Future work includes exploring joint learning of entities and events and applying GCNs to other information extraction tasks like Relation Extraction.
Summary updated