AI text simplification is transforming the way we process complex online information. While expert language is essential in fields like health, law, and finance, it often makes important details hard to understand. AI text simplification can be done with the tools like Gemini, you can instantly simplify dense text, making it easier for everyone to access and comprehend crucial information.
Today, Google Research presents “LLM-based Text Simplification and its Effect on User Comprehension and Cognitive Load,” detailing our new system built with Gemini models. Our focus is on minimally-lossy simplification: making text clearer without losing its original meaning, details, or nuances. This differs from summarizing (which cuts information) or explaining (which adds information). We’re also excited to launch this technology as a new “Simplify” feature in the Google app for iOS.
Achieving this requires AI models to accurately rephrase complex ideas without errors or omissions. The simplified text needs to boost understanding without compromising the original information’s integrity.
This work delivers two key advancements. First, we’ve created a novel system with automated evaluation and iterative prompt refinement. This allows Gemini models to rapidly discover the best way to simplify text while maintaining accuracy – something impossible to do manually at this scale. Second, a large-scale study proves that our text simplification significantly improves how well people understand the text and reduces how hard they have to work to do so.
Gemini-Powered Automatic Evaluation and Prompt Refinement
Our system uses Gemini models to automatically assess simplification quality and refine the instructions (prompts) given to the simplification model. Crafting prompts for subtle simplification is tough, requiring a balance between readability and accuracy. Our automated system solves this by allowing extensive experimentation to find the most effective prompts.
- Automated Evaluation: Manual assessment is too slow for quick improvements. Our system has two new evaluation methods:
- Readability Assessment: Instead of basic tools like Flesch-Kincaid, we use a Gemini prompt (refined against human judgment) to score text readability on a 1-10 scale, providing a more accurate measure of how easy text is to understand. We found this LLM-based assessment better matches human opinions.
- Fidelity Assessment: Preserving meaning is crucial. Using Gemini 1.5 Pro, we map claims from the original text to the simplified version. This pinpoints errors like lost, added, or distorted information, giving us a detailed measure of how faithful the simplification is to the original.
- Iterative Prompt Refinement: The quality of the simplified text (generated by Gemini 1.5 Flash) depends heavily on the initial instructions. We’ve automated the prompt improvement process. Another Gemini 1.5 Pro model analyzes the performance of the current prompt (based on readability and accuracy scores) and suggests better prompts for the next try. This creates a powerful cycle where the AI system learns to improve its own instructions, reducing the need for manual work and leading to much better simplification strategies. This process ran for 824 cycles until performance leveled off.
This automated loop, where one AI evaluates another’s output and refines its instructions based on performance and detailed errors, is a major innovation. It moves beyond slow manual work, allowing the system to independently discover effective ways to simplify complex text over many iterations.
Measuring Impact: A Large-Scale Randomized Study
To see how well this simplification works in the real world, we conducted a large study.
- Study Design: We recruited 4,563 participants with varying levels of topic knowledge. We used 31 diverse, real-world texts from areas like medicine, law, and finance. Participants were randomly assigned to read either the original text, the simplified version, or both. We also tested if simplification helped with remembering information, both when participants could look back at the text and when they couldn’t. We measured understanding with multiple-choice questions, self-reported confidence, and mental effort using a simplified NASA Task Load Index.
Results
With nearly 50,000 answers, our study clearly shows the value of simplification.
- Quantitative Findings: Participants reading the simplified text scored 4% higher on average in the comprehension questions compared to those reading the original. The biggest improvement (15%) was seen in very complex medical texts. We also saw significant gains in finance (6%), law (4%), and technical fields (4%). These improvements held true even when participants couldn’t refer back to the text. Participants also reported feeling more confident (a 0.24 increase on a -2 to 2 scale) and found the task easier (a 0.33 improvement on a -2 to 2 scale) with simplified text.
- Qualitative Insights: Looking at examples where simplification greatly improved understanding (by 38% for one medical text) shows how it helps. By defining jargon, breaking down long sentences, and clarifying relationships, simplification makes complex ideas much easier to grasp.
Simplification was especially helpful for texts that were initially hard to understand. More examples can be found in our research paper.
- Study Limitations: While large, our study has limitations. Our participants were recruited online and might not perfectly represent all users trying to understand complex information. While our system aims for accuracy, AI can still make mistakes, so ongoing monitoring is needed. Also, multiple-choice questions don’t fully measure deep understanding.
Available as a New Simplify Feature
Starting today, this technology is available as a “Simplify” feature in the Google app for iOS. Users can select any complex text on a webpage and tap the “Simplify” icon to instantly see a simpler version without leaving the page. This makes it easier to understand new or complex topics encountered online.
Conclusion
We’ve developed and rigorously tested an automated system using Gemini that learns to simplify text while keeping the original meaning intact. By significantly improving understanding and reducing mental effort for complex information, this technology empowers users across important areas.
Acknowledgements
This work was a team effort involving many individuals from different fields. Key contributors are listed.
Option 2: More User-Focused and Engaging
Ever feel lost in a sea of complicated online text? Whether it’s trying to understand a medical diagnosis, legal jargon, or financial reports, sometimes the language itself becomes the biggest hurdle. Imagine a world where you could instantly make that complex text clear and understandable.
That’s the idea behind our new research at Google, detailed in our paper “LLM-based Text Simplification and its Effect on User Comprehension and Cognitive Load.” We’ve built a system using the power of Gemini AI models to do just that: simplify text without losing its crucial meaning. Think of it as making complex information accessible, not just shorter (like a summary) or adding extra explanations. And we’re thrilled to announce that this technology is now available as a handy “Simplify” feature in the Google app for iOS!
Our goal is for the simplified text to help you truly understand challenging material, without missing any important details or accidentally changing the original message.
To achieve this, we’ve made two key breakthroughs. First, we’ve created a smart system that automatically figures out the best way to simplify text while keeping it accurate. It’s like having an AI constantly learning and improving how to make things clearer. Second, we put this to the test with thousands of people, and the results are clear: simplified text significantly boosts understanding and makes it easier to process information.
How Our Gemini-Powered System Works
Our system uses Gemini AI not just to simplify text, but also to evaluate how good the simplification is and to automatically improve the simplification process itself. Finding the right “instructions” (prompts) for the AI to simplify accurately is tricky. That’s why we built an automated system that can try out many different approaches to find the best ones.
- Smart Evaluation: We needed a way to quickly and accurately judge how well the AI was simplifying. So, we developed two new ways to evaluate:
- Readability Score: Instead of relying on old-fashioned readability tests, we trained a Gemini AI to score how easy text is to understand on a scale of 1 to 10. We found this AI-powered score to be much better at matching human opinions.
- Accuracy Check: We also use Gemini 1.5 Pro to compare the original text with the simplified version, looking for any lost, added, or distorted information. This gives us a detailed understanding of how well the meaning is preserved.
- Automatic Improvement: The quality of the simplified text depends on the initial instructions we give the AI. So, we built a system where one Gemini AI analyzes how well the simplification is working (based on the readability and accuracy scores) and then suggests better instructions for the AI to try next. This creates a continuous loop of improvement, allowing the AI to learn the best ways to simplify complex text without constant human intervention. This process ran hundreds of times!
This innovative approach, where AI evaluates and improves its own simplification strategies, is a big step forward. It allows us to go beyond manual tweaking and discover truly effective ways to make complex information accessible.
Does It Actually Help? Our Large-Scale Study Says Yes!
To see if our simplified text really makes a difference, we conducted a large study with over 4,500 participants. We used real-world examples of complex text from various fields and compared how well people understood the original versus the simplified versions. We also measured their confidence and how much mental effort they felt they put in.
The Results:
With nearly 50,000 questions answered, the results were clear: simplification works!
- People who read the simplified text scored an average of 4% higher on comprehension questions. The biggest improvement was in very technical medical texts, where scores jumped by 15%! We also saw significant gains in finance, law, and technical fields. Importantly, these improvements held even when people couldn’t look back at the text, suggesting better understanding and retention.
- Participants felt more confident in their answers and found the task easier when reading the simplified text.
- Looking at specific examples showed how simplification helps by explaining jargon, breaking down long sentences, and clarifying complex ideas. For instance, simplifying one medical text led to a 38% increase in correct answers!
While our study was large, it’s not perfect. Our participants were recruited online, and we relied on multiple-choice questions to measure understanding. However, the results strongly suggest that simplification is a valuable tool.
Try It Yourself: The New Simplify Feature
Starting today, you can experience this technology firsthand! If you have the Google app on your iPhone, simply select any complex text on a webpage and tap the new “Simplify” icon. You’ll instantly see a clearer version, right there on the page. This makes learning about new or difficult topics online much easier.
The Bottom Line
We’ve created and thoroughly tested a system using Gemini AI that learns to simplify text while keeping it accurate. By making complex information easier to understand, this technology has the potential to empower people in many important areas of their lives.
Thank You
This project was a collaborative effort involving many talented individuals. We’d like to thank the key contributors listed here.
Choose the option that best suits the tone and focus you want for your rewritten blog post. Both options aim to be clearer and more engaging than the original while still accurately conveying the key information.
Refferences