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

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computational Linguistics

Background:

• Large language models (LLMs) excel at abstractive summarization but often generate unsupported or hallucinated content.
• Ensuring factual accuracy and source faithfulness remains a challenge in automated text summarization.

Purpose of the Study:

• To introduce a biologically inspired bird-flocking framework as a preprocessing step for LLM-based text summarization.
• To enhance factual accuracy and reduce hallucinations in generated summaries by grounding LLMs in source text.

Main Methods:

• A bird-flocking-inspired algorithm identifies salient sentences using contextual information, sentence position, and thematic relevance.
• The framework incorporates unified stop-word control, collocation recognition, attention mechanisms, and unsupervised Flock-by-Leader text clustering.
• This method generates a structured intermediate representation of the source document for LLM input.

Main Results:

• The framework consistently produces concise and factually correct summaries, outperforming a major LLM baseline.
• Demonstrated improvements in ROUGE-1 (7.28%), ROUGE-L (6.19%), and entity coverage (45.28%) across over 9,000 documents.
• Significantly reduced model hallucination by ensuring summaries are exclusively grounded in the original text.

Conclusions:

• The bio-inspired bird-flocking framework effectively mitigates LLM hallucinations in text summarization.
• This approach serves as an efficient preprocessing step, complementing both conventional and generative AI summarization methods.
• The framework enhances summary consistency, diversity, and factual integrity, particularly for long-form documents.