Hierarchical Multi-Agent Extractive Framework for Long-Document Mind Map Generation
Abstract
A mind map is a hierarchical visual representation of a central idea and its associated concepts. It is designed to facilitate the understanding of complex information. Transforming long texts into structured graphs is a complex task in natural language processing. Existing approaches often rely on construct-ing dense relation matrices at the sentence level or
on the one-pass generation with Large language model (LLM), both of which suffer from severe computational limitations and degradation in reasoning as the context length increases. To overcome these limitations, we propose a new text-to-mind map pipeline driven by a hierarchical multi-agent framework. Our approach first uses adaptive semantic segmentation with iterative merging to partition long documents into coherent segments. We then introduce a three step multi-agent
pipeline: a Core Identifier extracts key plot events via a map reduce strategy, a Node Selector dynamically filters out irrelevant passages using segmentation and filtering logic, a Summarizer uses recursive batch
processing to synthesize the final text while adhering to token constraints. Evaluation on six large and varied datasets demonstrates the efficiency and robustness of our method. The proposed framework preserves essential semantic equivalence, achieving a BERTS F1 score of up to 0.65, thus generating visually navigable mind maps from large texts. The code is available at https://github.com/Noorius/Text-to-MindMap.
on the one-pass generation with Large language model (LLM), both of which suffer from severe computational limitations and degradation in reasoning as the context length increases. To overcome these limitations, we propose a new text-to-mind map pipeline driven by a hierarchical multi-agent framework. Our approach first uses adaptive semantic segmentation with iterative merging to partition long documents into coherent segments. We then introduce a three step multi-agent
pipeline: a Core Identifier extracts key plot events via a map reduce strategy, a Node Selector dynamically filters out irrelevant passages using segmentation and filtering logic, a Summarizer uses recursive batch
processing to synthesize the final text while adhering to token constraints. Evaluation on six large and varied datasets demonstrates the efficiency and robustness of our method. The proposed framework preserves essential semantic equivalence, achieving a BERTS F1 score of up to 0.65, thus generating visually navigable mind maps from large texts. The code is available at https://github.com/Noorius/Text-to-MindMap.
Keywords
Large language models, mind map, text summarization, multi-agent framework, artificial intelligence