A bird-inspired artificial intelligence framework for advanced large text summarization

TL;DR

This paper introduces a bird-flocking inspired AI framework that improves text summarization by reducing hallucinations and increasing factual accuracy.

Contribution

A novel bio-inspired bird-flocking framework for preprocessing text to enhance LLM-based summarization with source faithfulness.

Findings

The framework outperforms a major LLM baseline with gains in ROUGE-1, ROUGE-L, and entity coverage.

It reduces hallucinations by grounding summaries in the original text using salient sentences.

The method is effective across 9,000 long-form documents in healthcare and energy.

Abstract

We introduce a biologically inspired bird-flocking experimental framework for text summarization that identifies the most salient sentences using contextual information, sentence position, and thematic relevance. The bird-flocking-inspired algorithm, combined with large language models (LLMs), generates summaries with greater factual accuracy. The algorithm ensures source faithfulness by preventing the generation of new, unsupported information, thereby mitigating the risk of model hallucination by grounding the summary exclusively in the original text. While large language models (LLMs) achieve remarkable fluency in abstractive summarization, they frequently hallucinate generating plausible but unsupported content. We introduce a bio-inspired bird-flocking framework that addresses this limitation by serving as a preprocessing step for LLM-based summarization. Our method identifies the most salient, source-faithful sentences using contextual information, sentence position, and thematic relevance, providing LLMs with factually grounded input that constrains generation to verified content. Experimental results show that our methodology consistently produces concise and factually correct summaries, as experimented with the commonly used quality measurement scores. The framework provides a mechanism for text summarization that incorporates unified stop-word control, collocation recognition with synonym expansion, attention combination with fallback, score normalization between global and local saliency, and an unsupervised learning bio-inspired Flock-by-Leader text clustering algorithm. These components contribute not only to improved consistency and diversity of the summary, but also to reduced hallucinations in text summarization. The algorithms and experimental framework proposed in this study serve as an efficient preprocessing step that complements both conventional and generative AI-based text summarization methods. The framework produces a well-structured intermediate representation of the source document, which is then provided to the LLM to generate the final summary. Across over 9,000 long-form documents in healthcare and energy, our framework consistently outperforms a major large language model baseline, with gains of 7.28% in ROUGE-1, 6.19% in ROUGE-L, and 45.28% in entity coverage.

Figures