Generative large language models are all-purpose text analytics engines: text-to-text learning is all your need

Area of Science:

• Clinical Natural Language Processing (NLP)
• Artificial Intelligence in Healthcare
• Large Language Models (LLMs)

Background:

• Clinical NLP tasks traditionally require specialized models.
• Generative LLMs offer a potential unified approach.
• Prompt tuning enables efficient adaptation of LLMs.

Purpose of the Study:

• To develop and evaluate a unified text-to-text learning architecture for major clinical NLP tasks.
• To leverage a generative LLM with prompt tuning for improved performance.
• To demonstrate the
• one model for all
• capability in clinical NLP.

Main Methods:

• Formulated 7 key clinical NLP tasks as text-to-text problems.
• Utilized GatorTronGPT, a unified generative clinical LLM (GPT-3 architecture, up to 20 billion parameters).
• Employed prompt tuning with soft prompts and a frozen LLM.

Main Results:

• Achieved state-of-the-art performance on 5 out of 7 clinical NLP tasks.
• Outperformed previous task-specific Transformer models in concept/relation extraction, concept normalization, abbreviation disambiguation, and natural language inference.
• Surpassed a prior prompt-based model (GatorTron-MRC) for clinical concept and relation extraction.

Conclusions:

• The unified generative LLM with prompt tuning is effective for diverse clinical NLP tasks.
• This approach offers a promising
• one model for all
• solution from training to deployment.
• Significant performance gains were observed compared to existing methods.