Dynamic Spectral Clustering for Multi-Document Summarization of Social Media Discourse

The growing volume of socially generated content—ranging from online discussions to news articles—poses increasing challenges for collective understanding and information synthesis. Multi-document summarization offers a means to distill key ideas from such large and redundant text collections, yet existing large language models struggle to integrate dispersed context and maintain factual consistency. This paper presents a modular summarization framework that integrates dynamic spectral clustering with structured abstractive rewriting to improve both interpretability and coherence. Our method first partitions related documents into semantically coherent clusters using a dynamic spectral algorithm inspired by community detection, then produces localized summary drafts with a lightweight sequence-to-sequence model, and finally refines these drafts through an abstractive rewriting stage. This structured design enables controllable and transparent summarization across heterogeneous sources. Experiments on benchmark datasets demonstrate that the proposed approach achieves competitive results relative to recent large-model systems while offering clearer interpretability and efficiency advantages for social computing applications.