AI in Swedish can transform the public sector

A key factor in this change is the fact that powerful computers were used in 2020 to read large volumes of texts from the National Library of Sweden, which has a collection of everything printed in Sweden since 1661: newspapers, magazines, fiction books, non-fiction books, papers, etc.

The enormous text input has created an artificial language brain that could be of benefit to the government, businesses, researchers and others.

– “Using this language brain, it would, for example, be possible to develop much better AI-based text categorisation systems much more quickly than before,” says Magnus Sahlgren, Head of Natural Language Processing at RISE, who is leading the work to develop a first AI-based language comprehension model for the Swedish government.

Adapted for the government

A digital language comprehension model adapted for incoming and outgoing government communication is being created in the “Language models for the Swedish Government” project. RISE is developing the model using expertise in both AI and language model algorithms. The Swedish Public Employment Service, the Swedish Tax Agency and the Swedish Agency for Economic and Regional Growth are collaborative partners, together with AI Sweden, LTU and Peltarion.

– “The project ambition is for every government agency to be able to make use of the best in Swedish AI technology to manage their information flows.”

The most common such flows in the public sector will be inventoried in February and specific solutions that can be used by government agencies will be developed throughout the rest of the year.

With the language model in place, AI can quickly be trained to sort and categorise e-mails, reports and documents and to ensure that the correct case officers have access to the right information at the right time.

– “E-mail, as an example, can be forwarded automatically to the correct departments or correct people or categorised in a certain way.”

The computer “understands” the context and the intentions

A well-developed AI understands information just as well as people do, in some cases even better.

– “For example, an e-mail concerning a notice of tax assessment for self-employed persons (F-skattsedel) no longer needs to include the term ‘F-skattsedel’ for the AI to immediately understand the context and know what to do with the information.”

Based on the language model, citizens should be able to enter questions and receive answers directly via a government web page, as opposed to being referred to pages containing further information.

– “An important point of the language comprehension model is that it will require much less training data for the AI to comprehend quickly, as there is already an underlying knowledge of language.

The language model also means that it is not especially important how citizens formulate what they say. Two people who want to know the same thing but who express themselves differently will both be understood and will receive the same message.

– “The AI understands the meaning behind individual expressions. This is the key difference compared to regular keyword matching.”

A large number of tests will be performed to ensure that citizens do not receive inappropriate or biased responses.

– “Our model has been trained using editorial text, not Flashback, but we still do not know what impact it will have when the AI uses the language in government agencies. There are also a number of other uncertainties.” The model has, for example, been trained using Standard Swedish. Will it work just as well in Norrbotten as in the rest of the country? This remains to be seen, but will be managed.