2025/01/07 by Fang, Dengzhao, Jipeng Qiang, Yi Zhu +8 · 1 citation
Computer Science · #Computation and Language (cs.CL) #FOS: Computer and information sciences #Natural Language Processing Techniques #Text Readability and Simplification #Topic Modeling
paper · pdf · doi:10.48550/arxiv.2501.03857
openalex publication_date 2025/01/07 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
Research on text simplification has primarily focused on lexical and\nsentence-level changes. Long document-level simplification (DS) is still\nrelatively unexplored. Large Language Models (LLMs), like ChatGPT, have\nexcelled in many natural language processing tasks. However, their performance\non DS tasks is unsatisfactory, as they often treat DS as merely document\nsummarization. For the DS task, the generated long sequences not only must\nmaintain consistency with the original document throughout, but complete\nmoderate simplification operations encompassing discourses, sentences, and\nword-level simplifications. Human editors employ a hierarchical complexity\nsimplification strategy to simplify documents. This study delves into\nsimulating this strategy through the utilization of a multi-stage collaboration\nusing LLMs. We propose a progressive simplification method (ProgDS) by\nhierarchically decomposing the task, including the discourse-level,\ntopic-level, and lexical-level simplification. Experimental results demonstrate\nthat ProgDS significantly outperforms existing smaller models or direct\nprompting with LLMs, advancing the state-of-the-art in the document\nsimplification task.\n