LINKEDLIST-BART: A DYNAMIC CHUNKING STRATEGY FOR EFFICIENT LONG-DOCUMENT SUMMARIZATION

ABSTRACT

We present LinkedList-BART, an improvement to the chunk-based BART summarizing technique that dynamically manages sentence chunks using a linked-list data structure. In contrast to static chunking techniques, our method allows for efficient O(1) chunk insertion, deletion, and reorganization by representing sentence fragments as nodes in a linked list. By allowing adaptive chunk splitting or merging at inference time, this flexibility avoids token-overflow problems while maintaining important cross-chunk context for abstractive summarization. We evaluate LinkedList-BART using both ROUGE and semantic metrics (BERTScore, BLEURT) on standard datasets (CNN/DailyMail, PubMed, ArXiv) and the Multi-News corpus. The results indicate that on long texts, a LinkedList-BART consistently outperforms the original chunked BART and competing baselines (PEGASUS, Longformer-LED, and BRIO), reducing the performance gap to heavier models but maintaining minimal inference overhead.

Keywords: LinkedList, Summarization, BART, Long-Document, Abstractive