Data Augmentation with Large Language Models: A Scaling Law-Guided Approach

Large language models (LLMs) significantly advance data augmentation research. However, existing approaches largely overlook two fundamental issues. First, empirical evidence demonstrating the alignment between LLM natural language (LLMNL) and human natural language (HNL) remains limited. Second, current methodologies often neglect the variability of LLMNL in comparison with HNL. To address the gap, we introduce a comprehensive scaling-law-based framework for examining the congruence between LLMNL and HNL. Through extensive experiments, we uncover a progression of findings: LLMNL fails to achieve congruence with HNL; there is a consistent discrepancy, with Mandelbrot exponents for LLMNL being approximately 0.2 lower than those of HNL; LLMNL exhibits reduced fractal complexity, corroborated our analysis to stylistic factors such as readability, sentiment, and semantics. Furthermore, we propose a new data augmentation approach for text classification, which leverages scaling laws to make decisions on LLM-generated texts. Extensive experiments under real-world scenarios demonstrate that our approach is competitive and robust, outperforming recent methods and consistently maintaining performance advantages across different LLMs and prompts.