AI for Easy-to-Read Language Now Available: Check and Translate

A text that no one understands doesn’t inform anyone. That may sound obvious, but it’s a real problem in the day-to-day operations of government agencies, schools, and social service organizations. Plain language solves this problem, but manual translation is time-consuming and costly. That’s exactly where our netidee project comes in. It’s now complete, and all the results are publicly available.

What's Available Now

Easy Language AI checks German texts against the rules of easy language and generates a simplified version. There are three ways to access the system.

The web tool at check. simple-language. ai no registration or installation required. Paste your text, check it, and you're done.
The browser extension for Chrome, Edge, and Firefox checks selected text on any website via a sidebar.
The API can be integrated into your own applications and content management systems. It is designed to run on your own servers.

All of the code, both sets of documentation, and the design system are available under open licenses. If you want to run the system yourself, you'll find everything you need in the developer documentation.

Why the system must measure first and then translate

We made the most important decision in the project contrary to our original plan. We wanted to train a language model directly on Easy Language. However, we lack the necessary data. There is no large, publicly available collection of high-quality texts in Easy Language that could be used to reliably train a model.

So we reversed the order. The system first evaluates how good a text is. Eighteen rule modules check sentence length, subordinate clauses, the passive voice, foreign words, compound words, negations, numbers, abbreviations, and more. Each finding is justified and assigned to a specific rule. This evaluation is then fed back to a language model, which refines the text over several iterations until the rules are met.

Here’s an example. For the sentence “Receipt of funding requires that the applicant organization fulfill the documentation requirements communicated in advance by the deadline and submit the relevant documents electronically,” the system flags 31 violations across 9 rule categories. It lists each one individually, such as “Unfamiliar word: ‘organization’; simpler: ‘group’” or “Sentence 1 is too complex, with 23 words.”

The detour took some time. But it is the reason why the results are verifiable and don't just sound plausible.

An explanatory video for anyone hearing about this for the first time

Not every question can be answered with an API. That's why a 45-second explanatory video was also created as part of the project: Easy-to-Understand Explanatory Video . It explains what plain language is and how to recognize it: short sentences, familiar words, and pictures when needed. No prior knowledge is required, and no product is shown. The video is available under a free license; you can use it in your own training sessions and on your websites.

Accessibility in Design, Not Just in the Final Product

A tool for accessible language that is itself difficult to use defeats its purpose. That is why the project has evolved into a complete design system: leichte-sprache. ai/design-system. html . It uses the Atkinson Hyperlegible font throughout, complies with WCAG 2.1 AA, and displays the original text and simplified text side by side. The Figma source files are publicly available. They also include the icons used to visualize plain language as vector graphics; third parties can use them directly.

Thanks to OKAY, Office for Plain Language

Plain language cannot be automated at a desk. Our partner, OKAY, the Plain Language Office, conducted a full-day workshop with the team led by Helga Mock. The practical examples discussed during that workshop now inform our selection and prioritization of the rules.

What's Next

The system is currently running on our development server and is available for public use through it. The GPU hardware intended for production use in an Austrian data center is currently hard to come by; an order was canceled by the vendor. As soon as the hardware arrives, we will migrate the system. This will not affect availability for users.

In terms of content, we are working on the quality of the translation, on additional and subject-specific rules, and on feedback mechanisms within the tool. Contributions from the community are expressly encouraged. The rule modules are loaded dynamically; a new rule is a new file and does not require any changes to the core.

Anyone who wants to move the project forward has three obvious starting points: developing custom rule modules for technical terminology, modules for content management systems such as Drupal, WordPress, or TYPO3, and adapting the architecture to other languages.