"We are starting to let the machines learn."

How did you come to open up to the topic of AI as a "normal web agency"?

Armin Berger: That was a lengthy process. For some time now, we have been working on the topic of semantics, i.e. the linking of individual, initially independent information units. The topic is not new, but until now there were hardly any practical application scenarios. I wanted to change that.

So we started to develop technologically alongside our traditional business - the creation of websites and the installation of content management systems. We initially developed our own storytelling format using semantic technologies. Our aim was not only to offer users a significantly improved search experience, but also to develop and tell stories at the same time.

A small shift then took place in this work. Previously, we as an agency had always focussed on the front end. However, we realised that the editorial team also had an application problem. We developed a tool for them that can be used to develop storylines based on the semantically annotated database.

That was our first step towards AI. It was very well received by both editors and users. That gave us courage. We then received funding for a small research project on the topic of semantic technologies - a joint project with technological partners that we had got to know through the Xinnovations network. The research partners were the German Research Centre for Artificial Intelligence (DFKI) and other SME partners from Berlin. At the time, our research focussed on automated text analysis and recommendation algorithms for storytelling.

As a result, we developed an extension for the TYPO3 content management system that can be used to automatically analyse texts and link them to external knowledge sources. In this sense, the machine acts as a kind of assistant for the editor, who checks the results for correctness and can accept them at the click of a mouse.

The system is already of impressive quality, but it gets even better with every application. The editors are really enthusiastic because it makes their work easier and more efficient. The response to the 3pc editorial assistant has been correspondingly positive.

This includes, for example, the automated tagging of images and videos, the generation of captions and text summaries or recommendation algorithms for augmented storytelling.

The background to this is that there will be more and more content in the future. The volume is virtually exploding. At the same time, there are more and more devices and reception situations: Mobile phones, notebooks, tablets, watches, glasses, audio, apps, etc. Editors not only have to get through vast amounts of content, but also prepare it for a wide variety of formats - a major challenge in times when hardly anyone can afford a well-staffed editorial team. With QURATOR, we want to solve precisely this problem.

In your experience, what potential does artificial intelligence have?

We are very cautious about the potential of AI. The dilemma is usually that you expect too much from the technology and are then disappointed that it can't do as much as you had hoped. So the normal reaction is to let AI be and wait until it can do more.

We take a different approach. We set about letting the machines learn. We accept that the quality of the content suggested by an AI varies. 60 per cent may be good, 30 per cent mediocre and 10 per cent really bad. We let the editors, the "curators", assess this - and the system learns accordingly and continuously improves. This is how we want to avoid the dilemma of expecting too much from AI.

How is the market reacting to your development?

The demand for such systems is - it feels - gigantic. You can offer a tool like this because it works. All those editors who moan every time they enter text because they end up having to keyword everything can now breathe a sigh of relief. The tool really is a well-functioning aid.

The next step is now automated image, video and audio recognition to make multimedia files searchable. Using so-called media fragments, individual passages in a video can then be displayed to match the search term, for example.

How far along are we with AI in Germany? Is it still rather embryonic - or are we already a little further along?

Some of the applications developed there are somewhat disappointing. The dilemma is that everyone involved scatters to the four winds after every funded project. There are many developments that are half-finished because the money has run out. Overall, there is a lack of continuity.

The second problem is translation: a lot of money is invested in science. The result is something potentially ground-breaking, but it is never applied in a meaningful way. We are familiar with the phenomenon of the fax: developed in Germany, it has never been applied there. The problem has been recognised by politicians and massive efforts are being made to change this. However, I am very optimistic that things will improve in the future.