Divide-Prompt-Refine: a Training-Free, Structure-Aware Framework for Biomedical Abstract Generation

Technical Summary: Divide-Prompt-Refine (DPR-BAG)

Problem Statement
Biomedical abstracts are critical for downstream Natural Language Processing (NLP) tasks, including information retrieval, biocuration, and knowledge discovery, due to their high information density and structured organization. However, a significant portion of biomedical literature lacks abstracts. As of April 2026, approximately 29% of PubMed articles were missing abstracts, a volume driven by the growth of publication types such as case reports, editorials, and letters. These missing abstracts create a bottleneck for NLP applications.

The task of Biomedical Abstract Generation (BAG) aims to automatically generate abstracts from full-text articles. Unlike standard document summarization, BAG requires adherence to scientific reporting conventions (structured presentation of methods, results, and conclusions) and the preservation of fine-grained biomedical entities and quantitative findings. Existing approaches face two primary challenges:

1. Extractive Bias and Fragmentation: Traditional extractive methods often produce fragmented text lacking natural flow. Even advanced models like LongT5 exhibit strong extractive bias when processing long documents, relying on verbatim copying rather than synthesis.
2. Fidelity Risks in Long Contexts: State-of-the-art models (e.g., GPT-4) suffer from hallucinations and information omission when extracting from non-decomposed scientific full texts. Furthermore, forcing models to process complex long texts often leads to a degradation in factual consistency.

Methodology: DPR-BAG Framework
The authors propose DPR-BAG (Divide, Prompt, and Refine for Biomedical Abstract Generation), a training-free, zero-shot framework designed to generate coherent, factually grounded abstracts. The framework operates via a modular pipeline:

1. Structure-Aware Decomposition (Divide):
   Instead of processing the full text as a single block, DPR-BAG decomposes the document into six rhetorical facets based on the BOMRC schema (Background, Objectives, Methods, Results, Conclusions) plus an "Others" category for unclassified content.

   • Splitting Strategy: The authors utilize LLM-SSC (Large Language Model-based Sequential Sentence Classification) to assign rhetorical labels to sentences. They employ a First Sentence Labeling (FS) strategy, where the label of a paragraph's first sentence serves as the global label for that paragraph. This approach outperforms naive length-based splitting (NS) and section-header normalization (SH) by providing fine-grained contextual anchoring.

2. Parallel Summarization (Prompt):
   Each resulting facet is independently summarized using a pre-trained, instruction-tuned LLM (specifically Llama-3.2:3B in the experiments) in a zero-shot manner.

   • Prompting Strategies: The study evaluates three prompting complexities:
     • Basic Concise (BC): Minimal instructions with coarse-grained focal points.
     • Detailed Instruction (DI): Modeled after JMIR submission guidelines, enforcing granular details (sample sizes, p-values).
     • Structural Instruction (SI): Uses Markdown formatting to enforce structural schemas.
   • Entity Guidance: The framework tests optional grounding extensions, including TR-UMLS (TextRank extraction linked to UMLS concepts) and Chain-of-Thought (CoT) reasoning.

3. Refinement (Refine):
   The individual facet summaries are concatenated and passed to a final LLM-based refinement stage. This step synthesizes the fragmented outputs into a unified abstract, ensuring global discourse coherence and stylistic consistency.

4. Validation and Fallback:
   A robustness mechanism is implemented where, if all six facet summaries fail, the system regroups content into three broader categories (Intro, Main Idea, Results & Conclusions). If this fails, a "Lead-300" heuristic (first 300 characters) is used to guarantee non-empty output.

Key Contributions

1. DPR-BAG Framework: A novel, training-free, structure-aware method for BAG that decomposes full-text articles into rhetorical facets, performs parallel summarization, and refines the output.
2. PMC-MAD Dataset: The release of a distribution-aligned dataset of 46,309 biomedical full-text publications (missing abstracts) specifically curated for the BAG task.
3. Comprehensive Evaluation: A systematic comparison of DPR-BAG against strong extractive and fine-tuned abstractive baselines (LED, LongT5), demonstrating superior performance in abstractiveness and factual consistency without task-specific fine-tuning.
4. Ablation Insights: A detailed analysis of prompting strategies and splitting methods, revealing counterintuitive findings regarding prompt complexity.

Experimental Results
Evaluated on the PMC-MAD dataset, DPR-BAG (using the Basic Concise prompt) demonstrated the following:

• Abstractiveness: DPR-BAG generated abstracts that were significantly more abstractive (higher n-gram novelty, lower density) than fine-tuned baselines, closely resembling human-written abstracts.
• Factuality: Despite high abstractiveness, DPR-BAG maintained superior factual consistency (measured by AlignScore, MiniCheck, and SummaC) compared to both zero-shot and fine-tuned baselines. Fine-tuned models often sacrificed factuality for abstractiveness.
• Splitting Strategy: The First Sentence Labeling (FS) strategy significantly outperformed Section Header (SH) normalization, confirming that fine-grained semantic boundaries are necessary for faithful generation.
• Prompt Complexity Findings: Increasing prompt complexity (DI and SI) or injecting explicit entity guidance (TR-UMLS, CoT) generally degraded factual alignment. The authors observed a "distraction effect" where complex instructions caused the model to struggle with formatting constraints at the expense of source-grounded facts. Notably, while CoT improved semantic alignment in some contexts, it often reduced factual overlap with the source.
• Model Scaling: Scaling up to larger models (Qwen2.5:7B/14B) improved semantic alignment but further increased abstractiveness at the cost of factual overlap, yielding no overall advantage over the 3B backbone.

Significance
The paper underscores the potential of training-free, structure-aware frameworks for scalable biomedical abstract generation in low-resource settings. The findings suggest that for long-document summarization in the biomedical domain, decomposing text by rhetorical structure and using controlled, concise prompting is more effective than relying on complex instructions or fine-tuning. The DPR-BAG framework offers a practical solution to the bottleneck of missing abstracts, enabling the utility of a vast corpus of biomedical literature for downstream NLP applications without the computational cost of training specialized models.