Source-linked AI summary
Zero-Shot Transfer Learning for Event Extraction
Lifu Huang, Heng Ji, Kyunghyun Cho, Clare R. Voss
TL;DR
Previous event extraction methods have limited portability to new event types because they rely on annotations for existing types. This paper models extraction as grounding, jointly embeds event mentions and ontology types in a shared semantic space, and transfers knowledge to unseen types. Without annotations for 23 test types, the approach performs comparably to a supervised LSTM trained with 500 annotated event mentions.
Problem
Existing event extraction methods depend on annotated event mentions and cannot readily handle new event types without costly additional annotation.
Method
A transferable neural architecture jointly maps structured event mentions and ontology-defined event types into a shared semantic space and ranks types by similarity.
Results
Without annotated mentions for 23 test event types, the transfer approach achieves performance comparable to an LSTM trained with 500 annotated event mentions.
Takeaways & Limitations
Existing event ontologies and annotations for a small set of seen types can support extraction of unseen types without additional annotation.
Takeaways & Limitations
The approach can confuse types within the same scenario and relies on argument-path and argument-concept semantics when role labels are uninformative.
Abstract
from arXiv · showhide
Most previous event extraction studies have relied heavily on features derived from annotated event mentions, thus cannot be applied to new event types without annotation effort. In this work, we take a fresh look at event extraction and model it as a grounding problem. We design a transferable neural architecture, mapping event mentions and types jointly into a shared semantic space using structural and compositional neural networks, where the type of each event mention can be determined by the closest of all candidate types . By leveraging (1)~available manual annotations for a small set of existing event types and (2)~existing event ontologies, our framework applies to new event types without requiring additional annotation. Experiments on both existing event types (e.g., ACE, ERE) and new event types (e.g., FrameNet) demonstrate the effectiveness of our approach. \textit{Without any manual annotations} for 23 new event types, our zero-shot framework achieved performance comparable to a state-of-the-art supervised model which is trained from the annotations of 500 event mentions.
1 Introduction
The paper reframes event extraction as grounding event mentions to structured event types, addressing poor portability and limited annotation coverage. Its transferable approach uses existing schemas and annotations to extend extraction to unseen types without additional annotation.
- Motivation: Traditional supervised event extraction treats event types and argument roles as atomic symbols, making new types require costly fresh annotation.The problem is especially severe for ontologies containing thousands of event types.
- Structured Representation: Rich event ontologies represent event types with predefined argument roles, while event mentions can be structured from triggers and candidate arguments.The paper cites FrameNet, VerbNet, PropBank, and OntoNotes as examples of such ontologies.
- Grounding Formulation: The paper models extraction as grounding by mapping each event mention to its semantically closest event type in the ontology.Structural similarities between mentions and types complement lexical trigger semantics.
- Method: A transferable neural architecture jointly maps event-mention and event-type structures into a shared semantic space and ranks candidate types by similarity.The framework transfers knowledge from annotated seen types to unseen types.
- Advantages: The mapping and ranking function is independent of event types, enabling transfer from existing types to new types without additional annotation.This design supports broader reuse of existing resources.
- Advantages: Existing event ontologies can expand extraction coverage from several dozen types to thousands of types.The broader ontology scope is presented as a second advantage of the proposed view.
2 Approach
The approach jointly represents event mentions and ontology types as structures, maps them into a shared semantic space, and ranks candidate types by similarity. It uses AMR-derived candidates, compositional representations, shared CNNs, and a modified training objective to support zero-shot event and argument extraction.
- Candidate identification: Event extraction uses AMR parsing to identify candidate triggers and arguments, with trigger typing and argument typing following the same overall pipeline.Candidate arguments are specified by a subset of event-related AMR relations.
- Prediction: After training, new event mentions are projected into the semantic space and assigned their closest event type; arguments are ranked against role structures using the same shared framework.For argument typing, the model compares a trigger–argument path with structures representing predefined roles.
- Structure construction: Mention structures contain AMR-based tuples, while type structures contain event types paired with predefined argument roles.The two structures use different tuple contents but are designed for joint representation.
- Structure construction: Matrices compose AMR-relation semantics with concept pairs, while a tensor models the implicit relation between event types and argument roles.The resulting tuple representations incorporate relational semantics before structure-level encoding.
- Joint embedding: A weight-sharing CNN converts mention and type tuple sequences into structural representations used to learn a ranking function in the shared semantic space.Inputs are zero-padded feature maps, and max-pooling generates sequence representations.
- Learning: The model minimizes a hinge ranking loss so each annotated type outranks negative ontology types, while negative Other mentions and a new loss strategy address overfitting to seen types.Similarity is measured with cosine similarity, and the modified objective handles mentions whose correct label is Other.
3 Experiments
The experiments evaluate zero-shot event classification across ACE and expanded ontologies, compare structural and lexical transfer, and examine annotation efficiency and error patterns.
- ACE Event Classification: The experiments use ACE event types as seen and unseen classes, with the remaining 23 types reserved for zero-shot testing.Training varies the number of popular seen types, while evaluation uses annotations from unseen types.
- Zero-Shot Classification: Structural similarity outperforms lexical similarity for trigger and argument classification, and transfer improves as more seen types are added.The comparison uses the WSD-Embedding baseline and Hit@K results.
- Transfer Across Event Types: The transfer model is evaluated on similar and distinct unseen ACE subtypes, including Justice subtypes with shared argument-role structures.This tests whether structural relatedness affects transfer performance.
- New Event Types: The expanded ontology combines 33 ACE types with 1,161 FrameNet frames to assess transfer to many non-ACE event types.Some ACE types are aligned with FrameNet frames in the combined ontology.
- Annotation Efficiency: Without annotated mentions for the 23 test types, the transfer approach performs comparably to LSTM trained on 3,000 sentences containing 500 annotated event mentions.The LSTM learning curve is constructed by successively adding portions of 905 annotations across ten folds.
- Error Analysis: Most trigger errors occur between types in the same scenario, while argument classification is weakened by uninformative roles such as Entity and Organization.The examples include confusion between Being-Born and Giving-Birth and reliance on argument-path and concept semantics.
- New Event Types: The approach discovers events absent from ACE, but type assignment can remain wrong even when the correct scenario is identified.The analysis indicates that trigger sense should also match the event-type definition.
- Impact of AMR: Fine-grained AMR relations, including location and instrument, are more informative for argument-role inference than core-role-only SRL output.The comparison retains only core AMR roles for the SRL-style condition.
4 Related Work
Related work includes supervised event extraction, schema-free event discovery, and zero-shot learning in other tasks. The paper distinguishes its approach by combining ontology-defined event structures with transfer across event types.
- Supervised Event Extraction: Previous event extraction methods largely use symbolic or distributional features learned from substantial annotated training data.These methods treat event types and argument roles as classification targets.
- Schema-Free Event Discovery: Schema-free information extraction methods infer event types and roles by clustering similar events, but depend heavily on information redundancy.The paper states that this dependence limits such methods on inputs containing only a few sentences.
- Zero-Shot Learning: Zero-shot learning has been applied to vision, fine-grained name tagging, and relation extraction, but event extraction typically has far fewer seen types.Popular event schemas such as ACE define only 33 event types.
5 Conclusions and Future Work
The paper frames event extraction as grounding and transfers knowledge from annotated seen types to unseen types using a transferable neural architecture. It reports comparable performance to state-of-the-art supervised models without annotations for unseen types.
- The approach models event extraction as a grounding problem rather than conventional supervised classification.
- A transferable neural architecture leverages human-constructed event schemas and annotations for a small set of seen types to extract unseen types.
- Without any annotation, the approach achieves comparable performance with state-of-the-art supervised models trained from a large amount of labeled data.