Long Text Processing Framework Based on LLMs and Retrieval-Augmented Generation

In recent years, there has been extensive scholarly investigation into the applications of Large Language Models (LLMs) across diverse downstream natural language processing (NLP) tasks. As a fundamental task in the NLP field, document generation aims to efficiently produce valuable reports as a human substitute, while error correction focuses on detecting typographical or grammatical errors in documents. However, LLMs face significant challenges in generating and proofreading Long-Form texts due to their context length limitation. To address the above issues, we have designed a framework for large models to process long texts. For long-text generation, firstly, an LLM is employed to create an article outline, then the content of each subsection is iteratively generated. Secondly, the generation of each subsection incorporates retrieval-augmented generation (RAG) technology to leverage the user's personal knowledge base. For long-text error correction, the document is segmented into smaller sections, then iteratively invoke the LLM to correct each segment sequentially. Finally, the experimental results demonstrate that the proposed approach exhibits significant potential for both long-text generation and error correction tasks.