This paper proposes LMTransplant, a novel text augmentation paradigm leveraging large-scale language models (LLMs). To overcome the limitations of existing text augmentation methods, which primarily focus on lexical-level transformations and thus lack the diversity of transformations while maintaining meaning, LMTransplant integrates the source text with the extended context generated by the LLM and then has the LLM regenerate the transformed text. This allows the LLM to leverage its inherent knowledge to generate more diverse and creative content-level transformations while preserving the core properties of the source text. It outperforms existing methods on a variety of text-related tasks and demonstrates excellent scalability as the augmented data size increases.
