Abstract
Traditional event extraction relies on fixed schemas and extensive manual labeling. This paper presents SS-VQ-VAE (Semi-Supervised Vector Quantized Variational Autoencoder), a framework that leverages a small set of labeled "seen" event types to discover and categorize "unseen" types from raw text. By combining contextual embeddings, vector quantization, and a variational autoencoder regularizer, the model achieves state-of-the-art performance in supervised detection while effectively inducing high-quality new event categories.
Introduction
Event Extraction (EE) typically operates under a "closed-world" assumption, requiring a predefined list of event types and thousands of manually annotated examples. This approach is both expensive and brittle, as models fail when encountering novel event types not present in the training data. To address this "schema bottleneck," the authors propose Semi-supervised New Event Type Induction. The goal is to identify "seen" types with existing annotations while automatically grouping other mentions into coherent "unseen" categories.
Methodology: The SS-VQ-VAE Framework
The proposed SS-VQ-VAE framework integrates three core components to bridge the gap between supervised detection and unsupervised induction.
- Contextual Embeddings (BERT): A pre-trained BERT encoder generates rich, context-aware vectors for candidate trigger words, capturing semantic nuances (e.g., distinguishing "fired" as a dismissal versus a weapon discharge).
- Vector Quantization (VQ): This component serves as the clustering mechanism. The model maintains a dictionary of discrete latent type representations. Seen types are anchored by labeled data, while unseen types are learned by grouping similar unannotated mentions around learned centroids.
- VAE as a Regularizer: To ensure the model learns generalizable features rather than just memorizing seen types, a Variational Autoencoder (VAE) is employed. It attempts to reconstruct the original trigger's features from its predicted type, forcing the latent space to capture the semantic essence of all events.
Experimental Setup
The model was evaluated using the ACE 2005 dataset. In the supervised setting, all 33 event types were used to compare against traditional methods. In the induction setting, the 10 most frequent types were treated as "seen," while the remaining 23 were treated as "unseen" to test the discovery of new categories.
Results and Discussion
SS-VQ-VAE demonstrates superior performance in both supervised and unsupervised tasks. In supervised event detection on ACE 2005, the model achieved an F1-score of 76.7%, surpassing previous state-of-the-art models like JMEE and DMBERT.
For new type induction, the model significantly outperformed baseline clustering methods. It achieved a Normalized Mutual Information (NMI) score of 40.88, a substantial improvement over the 8.93 achieved by a standard BERT-based K-Means approach.
Qualitative analysis via t-SNE visualization confirms that the model induces high-quality clusters for unseen types. Mentions of events such as Sentence, Marry, and Convict form distinct, tight clusters despite the absence of explicit labels during training.
Conclusion and Future Work
The SS-VQ-VAE framework effectively addresses the schema bottleneck by leveraging seen types as anchors for discovering new event categories. While the model excels at type induction, limitations remain in trigger identification heuristics and the lack of automatic argument role induction. Future research will focus on extending this framework to induce event arguments and dynamically determining the number of unseen types in open-world scenarios.
Summary updated