Latent Graph Induction Networks and Dependency Graph Networks for Events Detection

The goal of event detection is to identify instances of various event types within text. In real-world scenarios, multiple events often coexist within the same sentence, making the extraction of these events more challenging than extracting a single event. While graph neural networks operating over dependency parsing trees have shown some capability in handling multi-event scenarios and improving event detection effectiveness, their improvement is limited. This limitation arises because dependency trees cannot automatically establish connections between trigger words and other key words, which are crucial for recognizing and classifying trigger words. Additionally, syntactic-based methods typically focus on the closest neighbors in the dependency graphs to aggregate information for the trigger candidate word, even though relevant words are often multi-hop away. In this paper, we combine the word dependency graphs with our automatically induced latent graph structure for event detection and multiple events detection. Furthermore, we propose two regularizers to enhance the representation of the dependency graphs and the induced latent graph structure. Experimental results demonstrate the effectiveness of our model for events detection.