Reproduce “CoDa: Constrained Generation based Data Augmentation for Low-Resource NLP” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Chandra Kiran Reddy Evuru, Sreyan Ghosh, Sonal Kumar, Ramaneswaran S, Utkarsh Tyagi, Dinesh Manocha

CoDa introduces a training-free data augmentation technique for low-resource natural language processing, leveraging instruction-tuned large language models guided by natural language constraints to produce diverse and consistent synthetic data. The method achieved F1 score improvements ranging from 0.12% to 7.19% across 11 datasets and three tasks, including sequence classification, named entity recognition, and question answering.

Key insights

  • Leveraging natural language constraints effectively guides instruction-tuned LLMs to generate high-quality, task-specific synthetic data without requiring fine-tuning, enhancing accessibility for low-resource NLP.
  • Explicit control mechanisms ensure a balance between diversity and consistency in augmentations, leading to more robust downstream models that generalize better.
  • Identifying and explicitly avoiding 'spurious concepts' during generation is crucial for creating less biased training data and improving model robustness against shortcuts.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.