Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
119 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Zero-Shot Transfer Learning for Event Extraction (1707.01066v1)

Published 4 Jul 2017 in cs.CL

Abstract: 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.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (4)
  1. Lifu Huang (92 papers)
  2. Heng Ji (266 papers)
  3. Kyunghyun Cho (292 papers)
  4. Clare R. Voss (14 papers)
Citations (198)

Summary

A Comprehensive Examination of Zero-Shot Transfer Learning for Event Extraction

The paper "Zero-Shot Transfer Learning for Event Extraction" presents an innovative approach to the challenge of modeling event extraction as a grounding problem, facilitating the transfer of knowledge from annotated to novel event types. The paper specifically addresses the limitations of traditional supervised methods that necessitate extensive annotations, rendering them impractical for new event categories.

The researchers propose a novel neural architecture that incorporates structural and compositional networks to align event mentions and event types into a shared semantic space. This mapping allows the framework to leverage minimal annotated examples of existing types and the hierarchical structures within established event ontologies, enabling the identification of new event types without further annotations.

Methodology Overview

The authors outline a transferable neural framework that conceptualizes event extraction as mapping to the closest event type within an ontology. This involves:

  • Structural Representation: Using Abstract Meaning Representation (AMR) to capture both event mentions and type structures. Event mentions are constructed from triggers and candidate arguments, while event types utilize predefined roles in ontologies such as FrameNet and ACE.
  • Zero-Shot Learning (ZSL): Borrowing from visual classification, the framework adopts ZSL by projecting event mentions and types into a multi-dimensional vector space. Event types lacking annotations are treated as unseen, and the framework learns a ranking function that generalizes from seen to unseen types.
  • Neural Architecture: The architecture employs Convolutional Neural Networks (CNNs) to process feature maps derived from the semantic structures of event mentions and types. The approach involves learning a composition function via matrices and tensors to integrate semantic relationships into word embeddings within event structures.

Experimental Validation

Through extensive experimentation, the authors evaluate the proposed method against both existing (e.g., ACE) and new event types (e.g., FrameNet). Results demonstrated that the zero-shot framework could achieve performance levels comparable to supervised models trained with annotations from 500 event mentions, without additional manual annotations for 23 new event types. Such results highlight the practicality of the framework in handling a diverse range of events, a capacity that traditional methods lack due to dependence on extensive labeled data.

Implications and Future Directions

The implications are twofold. Practically, the proposed system offers a cost-effective solution for broadening the scope of event extraction systems, critical for domains with rapid evolution of event types. Theoretically, it expands the application of zero-shot learning beyond conventional domains, showcasing its adaptability to text-based tasks where event ontologies can provide rich structures.

In future work, the authors propose integrating event definitions and richer contextual information to further enhance performance. This incorporation could potentially mitigate some observed deficiencies, such as misclassification within closely related scenarios, by enriching the semantic context of events.

The research underscores the potential of transfer learning techniques for enhancing the adaptability and scalability of event extraction systems, setting a foundation for continued advancements in natural language processing frameworks.