Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Targeted Augmentation for Low-Resource Event Extraction

Sijia Wang, Lifu Huang

TL;DR

This work addresses the scarcity of labeled data in event extraction by proposing a targeted augmentation framework (Talor-EE) that, together with back-validation, augments data with diverse, polarity-aware, accurate, and coherent event mentions. It combines dependent-context retrieval from external corpora, targeted generation to enrich event structures, negative augmentation to capture non-events, a back-validation loop with entailment and coherence checks, and a robust fine-tuning regime using a DEGREE-based generative EE model. Extensive experiments on ACE05-E and ERE under zero- and few-shot settings show consistent improvements across multiple backbones and generation agents, with diversity—measured by richer argument-role coverage—driving the largest gains. Limitations include potential over-generation of non-events and challenges in fine-grained argument-role distinctions, suggesting avenues for further refinement and broader application.

Abstract

Addressing the challenge of low-resource information extraction remains an ongoing issue due to the inherent information scarcity within limited training examples. Existing data augmentation methods, considered potential solutions, struggle to strike a balance between weak augmentation (e.g., synonym augmentation) and drastic augmentation (e.g., conditional generation without proper guidance). This paper introduces a novel paradigm that employs targeted augmentation and back validation to produce augmented examples with enhanced diversity, polarity, accuracy, and coherence. Extensive experimental results demonstrate the effectiveness of the proposed paradigm. Furthermore, identified limitations are discussed, shedding light on areas for future improvement.

Targeted Augmentation for Low-Resource Event Extraction

TL;DR

This work addresses the scarcity of labeled data in event extraction by proposing a targeted augmentation framework (Talor-EE) that, together with back-validation, augments data with diverse, polarity-aware, accurate, and coherent event mentions. It combines dependent-context retrieval from external corpora, targeted generation to enrich event structures, negative augmentation to capture non-events, a back-validation loop with entailment and coherence checks, and a robust fine-tuning regime using a DEGREE-based generative EE model. Extensive experiments on ACE05-E and ERE under zero- and few-shot settings show consistent improvements across multiple backbones and generation agents, with diversity—measured by richer argument-role coverage—driving the largest gains. Limitations include potential over-generation of non-events and challenges in fine-grained argument-role distinctions, suggesting avenues for further refinement and broader application.

Abstract

Addressing the challenge of low-resource information extraction remains an ongoing issue due to the inherent information scarcity within limited training examples. Existing data augmentation methods, considered potential solutions, struggle to strike a balance between weak augmentation (e.g., synonym augmentation) and drastic augmentation (e.g., conditional generation without proper guidance). This paper introduces a novel paradigm that employs targeted augmentation and back validation to produce augmented examples with enhanced diversity, polarity, accuracy, and coherence. Extensive experimental results demonstrate the effectiveness of the proposed paradigm. Furthermore, identified limitations are discussed, shedding light on areas for future improvement.
Paper Structure (30 sections, 5 figures, 8 tables, 1 algorithm)

This paper contains 30 sections, 5 figures, 8 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Low-resource Event Extraction
  4. Data Augmentation
  5. Controlled Text Generation
  6. Learning with noisy labels
  7. Model
  8. Problem Formulation
  9. Targeted Augmentation [Diversity]
  10. Dependent Context Retrieval
  11. Targeted Generation
  12. Negative Augmentation [Polarity]
  13. Back-Validation
  14. Event Mention Accuracy Verification [Accuracy]
  15. Event Mention Coherence Verification [Coherence]
  16. ...and 15 more sections

Figures (5)

  • Figure 1: Talor-EE framework overview.
  • Figure 2: Event mention accuracy verification module.
  • Figure 3: Event mention coherence verification module.
  • Figure 4: Experimental results on ACE05-E. (a-b) are visualizations for Common 5, and (c-d) for Common 10.
  • Figure 5: Experimental results on ERE. (a-b) are visualizations for Common 5, and (c-d) for Common 10.