Applying AI in Schibsted
At Schibsted, we experiment and work with AI in all our business areas. These are some examples of applications that have improved the way we work and our products.
Applying AI in Schibsted
At Schibsted, we experiment and work with AI in all our business areas. These are some examples of applications that have improved the way we work and our products.
Video-subtitles
Over the last few years, the demand for subtitled videos has increased due to accessibility needs and because more and more users watch videos without sound. But the manual timing and writing of subtitles take a lot of time. With the help of the OpenAI-model Whisper, video editors can now upload videos to a service built by the Aftonbladet TV Operations team and it will automatically generate subtitles. These subtitles can then be used in Adobe Premiere.
“The subtitles still need to be edited in Premiere a bit to fix minor errors, but the time gained is enormous, which creates more time for other tasks in the newsroom,” says Vasilios Hatciliamis, Head of TV Operations at Aftonbladet.
Language model
In a small cabin during summer vacation 2023, a Schibsted LLM was born. Simen Eide and Anders Haarr from AI foundations in Schibsted started training a model with Schibsted content to create SEO optimised headlines for Schibsted newspapers. It turned out to be five times as good at the job than ChatGPT and other open-source models. So far it’s been tried out on VG but the internal interest to use it is big. The ambition is also to implement it on other products in Schibsted, like marketplace brands, but exactly where this will go is not decided.
“It’s really cool that we are able to work on a project like this, without a final goal,” says Simen. ”I think that’s thanks to the culture of innovation that we have in Schibsted.”
AI helps engineers in Schibsted to write code.
AI helps engineers in Schibsted to write code.
AI helps with writing code
Hundreds of software engineers in Schibsted now use artificial intelligence to help them write code. “My productivity has grown at least 15%,” estimates one of them, Pedro Goncalves.
Good software engineering requires lots of creativity as well as superb skills in solving problems. But the daily work is also full of tedious and repetitive tasks. There are tests and failures as new code is produced – and it all takes time and energy.
That’s why, after a pilot project, Schibsted decided to let all its software engineers use the AI tool GitHub Copilot in their daily work. GitHub Copilot is like a ChatGPT for programmers, it just suggests software code instead of normal text.
After only a few weeks, 34,000 lines of code had been accepted by the software engineers. (By comparison: some estimate that an average programmer writes about 25,000 lines of code in a year.) And nearly 400 programmers have already started using it.
Indexing offers at Prisjakt
At Prisjakt, a price comparison service within the Schibsted family, machine learning (ML) plays an integral role in their ever-so-important product matching system.
The system utilises ML algorithms, such as neural networks, natural language processing (NLP) and computer vision methods, that support the categorisation of items from shops and matches the items to Prisjakt’s database of products, called “product matching.” This is a redundant system that evaluates several inputs, such as text and images from the product descriptions, along with the price.
There are almost 300 million items to process in the system and millions are matched every day, effectively reducing the need for manual labelling for just a small group of people.
Manual labelling combined with the creation of new products helps to continuously improve the performance of the ML models, which results in more product matches. For the users on Prisjakt’s site, this results in a seamless experience as they search for and compare products, ultimately enhancing the shopping experience and driving more traffic to the shops.
Overall, Prisjakt’s well integrated ML solutions for automated product matching serve as a competitive advantage, as well as being the backend for the price comparison service.
A feature built with GPT-4, creates short summaries of news articles.
A feature built with GPT-4, creates short summaries of news articles.
AI-generated summaries
On VG, Aftonbladet and Aftenposten, readers can get short summaries of news articles. This is a feature built with GPT-4, created by a cross-brand team in Schibsted. When an article is ready, the journalist simply toggles on the functionality in the content management system, and a summary is generated. The journalists can then review it and hit ‘publish’ when it’s ready.
To make sure that mistakes do not slip through the cracks – there is an extra safety mechanism. The team behind it has asked GPT-4 to double-check that the text and the summary are aligned before it’s published, using text classification. And the readers like it. The overall click-through rate on VG’s summaries is 19%, and for young readers, it’s 27%. You might think that there is a risk that these readers don’t read the whole article – but it turns out that they often do. They simply use the summary as an introduction.
Transforming sound to text
When software engineer Johannes Andersen had lunch with a VG colleague who was moaning about having to transcribe an interview, he was sympathetic but didn’t think more about it until he saw a forum post about the OpenAI model Whisper. Then he tried the model out on a hack day and it turned out to work very well.
As it became clear how much time the tool could save, Johannes Andersen and his team invested three weeks to work on an application. They’ve built an interface and an app where journalists can upload their sound files and turn them into texts. At the time of writing, a few months after launch – the web application has saved 13,433 hours for journalists who can now use that time to do other things.
You can use it for several languages and the word error rate is 9.5% for Norwegian and only 4.2% for English. With the app, the journalists can work locally on their computers, so they don’t have to share any data.
A way to listen to all articles
In the last Schibsted Future Report, we told the story about Aftenposten’s synthetic voice. In cooperation with the company BeyondWords, a team from Schibsted’s subscription newspapers trained an artificial voice, using sentences from published articles that were recorded by podcast host Anne Lindholm. Since then, the project has grown. Now you can listen to most articles published on Aftenposten.no. And soon there will also be playlists with recommendations for articles to listen to.
At Aftenposten Junior, where the project started, you can choose between nine different languages, including Arabic, Ukrainian and Somali. And more newspapers in Schibsted are on their way to implementing the technology. But this project is not only about convenience. It’s also a question of giving everyone access to the same information.
“For instance, we learned from teachers that 92% of them have students who struggle to read in their classrooms,” says product manager Lena Beate Hamborg Pedersen.