Medical Text Simplification Using Reinforcement Learning (TESLEA): Deep Learning-Based Text Simplification Approach.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Artificial Intelligence

Background:

• Medical abstracts are publicly available but use complex vocabulary, hindering public understanding.
• Simplifying medical text is crucial for disseminating research findings to a broader audience.

Purpose of the Study:

• To develop a deep learning-based text simplification (TS) approach for converting complex medical text into simpler versions.
• To ensure the generated simplified text maintains high quality and accuracy.

Main Methods:

• A transformer-based language model combined with reinforcement learning was employed for TS.
• Optimized rewards included Flesch-Kincaid, relevance, and lexical simplicity to simplify jargon-dense medical paragraphs.
• The model was trained on 3568 complex-simple medical paragraph pairs and evaluated on 480 pairs.

Main Results:

• The proposed method achieved a Flesch-Kincaid score of 11.84, outperforming previous baselines.
• Comparable performance was observed with other baselines using ROUGE-1 (0.39), ROUGE-2 (0.11), and SARI (0.40) scores.
• Human evaluation indicated over 70% agreement on fluency, coherence, and adequacy.

Conclusions:

• A novel medical TS approach using reinforcement learning was successfully developed.
• This approach accurately simplifies complex medical paragraphs, enhancing readability and accessibility.
• The TS method can automate the generation of simplified medical text, broadening access to biomedical research.