Minimally-Lossy Text Simplification with Gemini: Making Complex Content Accessible

May 6, 2025

Diego Ardila, Software Engineer, and Sujay Kakarmath, Product Manager, Google Research

The digital age has democratized access to information, but it's often still locked behind complex language and jargon. This can be a significant barrier for users trying to understand critical content like health information, legal documents, or financial details. To address this, Google Research introduces a new system using Gemini models specifically designed for minimally-lossy (high-fidelity) text simplification. The goal is to enhance clarity while preserving the original meaning, detail, and nuance-distinct from summarization, which can drop information, or explanation, which might add it.

Key Technical Changes

• Model Architecture: The system leverages Gemini models, which are large language models (LLMs) fine-tuned for specific tasks. These models are trained on a diverse dataset to understand and rephrase complex content accurately.
• Evaluation Loop: A novel automated evaluation and iterative prompt refinement loop ensures the output remains high-fidelity. This loop continuously tests the model's output against the original text, refining prompts to minimize loss of information.

Implementation Details

• Data Collection: The training data includes a mix of complex texts from various domains (e.g., medical, legal, financial) and their simplified versions. This diverse dataset helps the model generalize across different types of content.
• Prompt Engineering: Initial prompts are designed to guide the model towards accurate simplification. The iterative refinement loop adjusts these prompts based on evaluation metrics, ensuring the output remains faithful to the original meaning.
• Evaluation Metrics: Key metrics include readability scores (e.g., Flesch-Kincaid Grade Level), semantic similarity (e.g., BERTScore), and user comprehension tests.

Use Case: Google App for iOS

The new "Simplify" feature in the Google app for iOS is a practical application of this research. Users can select complex text, and the app will generate a simplified version on the fly. This feature aims to make information more accessible to a broader audience without compromising accuracy.

Benchmarks and Results

• Readability Improvements: The system significantly improves readability scores while maintaining high semantic similarity. For example, Flesch-Kincaid Grade Level drops from an average of 12 (high school level) to 6 (middle school level).
• User Comprehension: User studies show that simplified text generated by the model leads to better comprehension and lower cognitive load compared to the original complex text.
• Preservation of Detail: The system excels at preserving key details, ensuring that no critical information is lost in the simplification process.

Why It Matters

For practitioners, this research highlights the potential of LLMs in making expert knowledge more accessible. By focusing on minimally-lossy text simplification, Google Research addresses a significant gap in current NLP tools. This can have far-reaching implications for fields like education, healthcare, and legal services, where clear communication is crucial.

Conclusion

The introduction of the Gemini-based text simplification system marks a significant step forward in making complex information more accessible. By combining advanced LLMs with a robust evaluation framework, Google Research has developed a tool that can help users understand challenging content without sacrificing accuracy or detail. This technology has the potential to empower a broader audience and bridge the gap between expert knowledge and everyday users.