Improving Low-Resource Text Generation with LLM Post-Editing

This technical report presents a novel approach to enhance low-resource text generation tasks using large language models (LLMs) through a programmer-interpreter framework. It addresses the limitations of existing post-editing methods that rely on smaller language models, which may not effectively generalize across domains. The proposed method combines the strengths of a programmer, which encodes task-specific knowledge, and an interpreter, which utilizes the domain generalization abilities of larger models like GPT-3.5 and GPT-4. The report details the iterative refinement process, where the initial output generated by the LLM is improved through specific edit instructions. Extensive experiments demonstrate the effectiveness of this approach, highlighting improvements in logical form-to-text conversion and low-resource machine translation. Additionally, the report elaborates on the framework's architecture, delineating the roles of the generator, programmer, and interpreter, while providing insights into the empirical studies validating its superior performance over conventional methods.