Learning word-level data augmentation for event detection in Nature Language Processing

In this study, we introduce Event AutoAugment (EAA), a novel adaptive data augmentation (DA) technique tailored for event detection in Natural Language Processing (NLP). EAA optimizes word-level editing operations using a reinforcement learning framework, where the augmentation strategies for each word are determined through a discrete search over a predefined set of actions, including deletions, synonyms replacement, and more. The efficacy of these strategies hinges on a carefully formulated reward function that assesses the impact of augmentations on the event detection model's performance. This reward function provides feedback on the suitability of applied augmentations, guiding the augmentation model to discover a jointly optimal policy that adaptively enhances data diversity and model robustness. Our framework is evaluated on two established benchmark datasets, ACE05 and MAVEN, which feature a wide variety of event types. To simulate low-resource conditions—a common challenge in real-world NLP applications—we employ stratified subsets of these datasets, using 1% to 30% of the original data size while preserving the distribution of event types. The adaptability and efficacy of EAA are evidenced by substantial improvements in model accuracy across these varying data sizes. For instance, on the ACE05 dataset, EAA achieves a 78% increase in the F1 score in the smallest data subset compared to the baseline without augmentation. Similarly, on the MAVEN dataset, it shows a 120% increase, highlighting EAA's capability to significantly enhance data diversity and robustness, outperforming both learnable and rule-based augmentation methods across different scales of data scarcity.