Dutch government aims to communicate in a way that is as accessible as possible to Dutch citizens. In some cases, this proves challenging. Some of the content to be communicated is complex or should be completely correct from a legal perspective. This implies the use of more complex language. In this blog series of 3, we will highlight a research project of a Dutch university and government agency. The aim of this research is to learn about the use of AI to support the accessibility of government communication.
The governments’ challenge
The Netherlands Enterprise Agency (RVO) helps entrepreneurs and organisations to invest, develop and expand their businesses and projects both in The Netherlands and abroad. RVO is a government agency which is part of the Dutch Ministry of Economic Affairs. Part of RVO’s communication with its clients is through letters. In order to meet different levels of reading skills, the aim is to write these in the B1 language level. This is the level that 95% of the Dutch population understands. To achieve this in practice is a challenge, because it is not always easy to decide which words and writing style exactly match the B1 level. And what to do with words that simply have to be in the letter, from a legal point of view?
A style guide offers writers support. The intended writing style is personal, with a short distance to the customer. The guide advises to use short sentences and simple words. “Simple” is described here as words with no more than three syllables. In addition, they should be concrete rather than abstract and be common in everyday language. These rules are clear, but they also make the writing of letters more difficult. Sometimes there are no alternative words to be found, or there is no insight into the commonality of a word in general.
How Large Language models may support
Language technology can help with this. On the one hand, it can analyze and highlight the more difficult parts of a letter. On the other hand, it can suggest alternatives. A well-functioning system can save a lot of time in delivering a clear and readable letter written by a human writer. New developments in the field of AI give hope in that regard.
Large Language Models (LLM’s) can generate fluent sentences that are often semantically coherent with what was written before. In addition, they can be steered towards writing in a certain style. With these capabilities, LLM’s can be deployed to reformulate existing sentences into simpler alternatives. There are some challenges, however. It is not certain beforehand which LLM is best at simplifying a letter in the domain of RVO. Within this project, we tried out different models that we could run ourselves in a safe environment. Finally, we experimented with different ways to offer our models examples of simplified RVO letters. In blog 2 we will go into the technical details of our experimentation – the models that we tried out and the considerations for using these models.
The universities’ research into the use of AI in communication
RVO started a collaboration with the Institute for Language Sciences (ILS) at Utrecht University this year. The aim is to explore the application of language technology for support in simplifying letters.The ILS is focused on a variety of research directions that deal with language and communication. One of them is to study whether a communicative act is performed in an effective way to convey a certain message to a certain audience. An outcome of this work is a tool called LiNT to assess text readability, which we will describe in more detail in the third blog of this series.
Within the ILS, a group of researchers applies AI methodology to language. This enables the analysis of a large number of texts, and automation or support of certain procedures. For the ILS, the project that RVO presented is an interesting applied case in a field that is developing. Part of the research has already been completed. This resulted in insights about the best approach and usability of the application. The next step will be to test the improved letters with RVO’s clients. This way, we can measure the actual improvement and impact on communication.
A solid step in improving government communication
Overall, the first results are promising. There are solid indications that language technology can support in the improvement of communication in the letters of RVO. The results also indicate that there is a wider applicability in accessibility of government communication.
In the second blog, we will talk about large language models and how they can be used to help simplify a text. We will also discuss the different ways in which this can be done and the disadvantages. In the third blog, we will address the question: what makes a text more readable? To this end, we dissect the textual elements that influence readability and describe research into a system for automatically estimating the readability of texts.
Co-author
This blog is co-authored by Florian Kunneman. Florian is Assistant Professor at the Institute for Language Sciences at Utrecht University. His research is aimed at supporting governmental communication by means of language technology. To this end, he studies online discussion fora to gain insight into societal trends, he improves on how conversational systems (chatbots) converse with different types of users and he works on automated text simplification.