EAG: A Three-Stage Biomedical Data-to-Text Generation Framework for Low-Resource Scenarios

This paper proposes the Enrich-Aggregate-Generate (EAG) three-stage framework, addressing the unique challenges of data-to-text generation tasks in the biomedical field, with a focus on resolving application issues of large language models in low-resource scenarios, aiming to enhance the accuracy, domain adaptability, and practicality of generated text.

Biomedical data-to-text generation is an important task that converts structured biomedical data (such as medical records, gene sequences, etc.) into readable text, applied in scenarios like medical report generation and scientific research assistance. However, this field faces three major challenges: 1. High text professionalism with a large number of technical terms; 2. Scarcity of high-quality annotated data and high acquisition costs; 3. Extremely high accuracy requirements for generated content—errors may lead to serious medical consequences.

EAG framework improves generation quality in low-resource scenarios through three stages:

Enrich Stage

• Structured data understanding: Parse data such as tables and graphs, extract key entities and attributes;
• External knowledge integration: Link to authoritative knowledge bases like UMLS and SNOMED CT to enrich semantics;
• Data synthesis and augmentation: Generate synthetic samples using rule templates, and augment existing data via techniques like back-translation.

Aggregate Stage

• Multi-source data fusion: Integrate multi-source information from electronic medical records, laboratory systems, etc., to build a unified view;
• Temporal information modeling: Capture temporal patterns and causal relationships of disease progression and treatment effects;
• Key information filtering: Filter information relevant to the generation target via attention mechanisms.

Generate Stage

• Domain-adaptive generation: Adapt to the biomedical domain via continued pre-training and instruction fine-tuning;
• Factual consistency constraints: Verify numerical accuracy and logical consistency;
• Controllable generation strategies: Support text generation in different styles (concise/detailed, professional/patient-friendly).

EAG is optimized for low-resource scenarios:

1. Efficient parameter fine-tuning: Use LoRA and Adapter techniques to train only a small number of parameters for domain adaptation;
2. Transfer learning: Quickly adapt to target tasks based on general or related biomedical pre-trained models;
3. Active learning: Intelligently select high-value samples for annotation to maximize annotation utility;
4. Multi-task joint training: Combine auxiliary tasks like entity recognition and relation extraction to improve main task performance.

Application scenarios of the EAG framework include:

• Clinical report generation: Automatically convert test results into standardized reports to reduce doctors' workload;
• Medical record summary generation: Extract key information from electronic medical records to generate concise summaries, supporting clinical decision-making;
• Scientific research data description: Convert experimental data into paper text to assist scientific writing;
• Patient education materials: Generate easy-to-understand content to help patients understand their health conditions.

The EAG project has been open-sourced on GitHub, with contributions including:

• Reproducibility guarantee: Provide complete code to facilitate verification of experimental results;
• Benchmark establishment: Serve as a benchmark method for biomedical data-to-text generation;
• Community collaboration: Attract global researchers to participate in improving and expanding applications;
• Educational resources: Provide practical references for learners in biomedical NLP.

The EAG framework provides a systematic solution for low-resource biomedical text generation through its three-stage architecture, emphasizing factual accuracy and domain adaptability. In the future, it can be combined with multimodal learning (integrating imaging and genomic data), reinforcement learning optimization, and interpretability research to further enhance the accuracy and reliability of the technology.