The Significance of Style Diversity in Annotation-Free Synthetic Data Generation
Researchers have developed a new framework for generating synthetic dialogue data without requiring human annotations, which are often scarce in rapidly evolving industrial settings. This method uses intent definitions and incorporates topic and style attributes to enhance data diversity, employing two novel stylization models, Univ and Exam, to create more human-like linguistic styles. An LLM-as-a-judge filtering process further refines data quality, achieving up to 93.3% of the performance of human-annotated data. The study highlights that style diversity is more crucial than topic diversity for synthetic data utility, and that integrating style attributes during generation is more effective than post-hoc adaptation. AI
IMPACT This research could significantly reduce the cost and time required to create training data for intent classification models, potentially accelerating AI development in data-scarce environments.