Reproduce “Data Generation Using Large Language Models for Text Classification: An Empirical Case Study” 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
Yinheng Li, Rogerio Bonatti, Sara Abdali, Justin Wagle, Kazuhito Koishida
Researchers at Microsoft Corporation empirically investigated the effectiveness of large language models in generating synthetic data for text classification, providing actionable recommendations for practitioners. The study demonstrates that mixing even a small amount of raw data with synthetic data is crucial for performance and that synthetic data is most impactful in low-resource settings, while also identifying potential biases introduced during generation.
Key insights
- Mixing even a small amount (e.g., 100 examples) of raw, human-labeled data with synthetic data is crucial for grounding the downstream model and achieving optimal performance.
- An LLM's direct performance on a task does not consistently correlate with the quality of data it generates for training other models.
- Synthetic data augmentation provides the most substantial performance improvements (3% to 26%) in low-resource scenarios, with diminishing returns as the raw data quantity increases.
- Subtle biases can be inadvertently introduced by LLMs through in-context examples, requiring vigilance and potential rephrasing of generated content to mitigate negative impacts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.