SemAug: Shaping the Future of Semantically-Enriched, Format-Specific Data Augmentation

Anonymous

Abstract: In the realm of artificial intelligence, the significance of high-quality data cannot be overstated, especially data that adheres to stringent formatting rules and structures. Addressing this need, our study introduces an advanced data augmentation method specifically designed for format-specific datasets. This method utilizes the capabilities of Large Language Models (LLMs) to generate data that not only meets the rigid formatting criteria but also maintains the integrity of the information. Central to our approach is the integration of specific format requirements into natural language prompts, which guides the LLMs to produce precisely formatted outputs. A salient feature of our approach is its self-evaluative mechanism, which autonomously assesses the semantic quality of the augmented data, distinguishing it from prior methodologies that require manual validation, thereby streamlining the augmentation process. Our research represents a pioneering step forward, enabling more efficient enhancement of datasets that demand exacting format adherence without the extensive resource investment typically associated with such tasks.

Paper Type: long

Research Area: Generation

Contribution Types: NLP engineering experiment

Languages Studied: English, Programming Languages