Situation : Ouest-France is one of the most iconic French newspapers. It has been around since 1899 and it still distributes 685,000 copies a day (making it the best selling paper in France). All its archives have been digitized (yes, that's millions of them) but too few people benefit from it.

Brief: How could we make this amazing content (archives) accessible to the greatest possible number of people, in the most user-friendly way?

Objectives:

- Building a brand new and innovative content experience

- Renewing the relation between Ouest-France and its readers

- Creating value out of the archives

Meet Archie, the first-ever AI archivist.

Archie is an AI that explores the archives of Ouest-France to find the articles from the past that allow people to understand today’s news a little better.

This idea was born out of data. Out of millions of pieces of amazingly insightful data that had been digitized and stored by Ouest-France but had never been made accessible to people. So we created Archie: a one of a kind content experience, data-driven, AI-fueled, deeply innovative.

The goal of Archie is to allow everyone to take a step back from the news and look at the bigger, historical picture. The user starts by chatting with Archie in the Google Assistant and chooses the news he or she is interested in. Then, Archie digs into the archives and finds the articles that are the most likely to give depth and context to this piece of news.

Once the user chooses a news article that interests her or him, Archie starts digging into 120 years worth of archives.

To make this possible, the archives were processed in several ways beforehand: visual recognition, spellchecking, identification of topics and lexical fields inside the articles. In the end, all of the archives were categorized in a graph database.

Here is how Archie does its job: it starts by doing a semantic analysis in order to identify the topics and lexical fields that can be found in the article.

Then, it uploads the article as a node into the archives graph database. This allows to compare the article with all the archives and find the ones with the highest similarity score.The articles selected by Archie are then greenlit by the editorial teams at Ouest-France.

Finally, when displaying the archives to the users, Archie customizes its output depending on the device: if the user is speaking to a Google Home, the content will be different to the one that will be shown on a smartphone.

To build this AI archivist, the digitized archives were processed using optical character recognition technology (OCR) in order to extract the text. Then a semantic algorithm was used to classify the archives into topics and match each news article with the most relevant archives. In the end, Archie submits its work to the editorial team of Ouest-France which decides whether to approve it or not.

One of the big challenges was to translate a human intuition (the fact that two articles might be related) into mathematical rules, to build an algorithm that could replicate the feeling of similarity and relevance that a real human archivist could feel when looking at two pieces of news. That means answering some quite abstract questions like: What is ""relevance""? What is ""resemblance""? What is ""meaning""?

That's why the creative and data-science teams worked hand in hand and iterated for months to try and produce the best possible semantic algorithm (and to translate meaning into ones and zeros).Thanks to this work, articles from the past are matched with ones from the present everyday. The algorithm works in real-time so as to always offer up-to-date content.

The development of Archie also implied to create a true experience inside the Google Assistant: Archie was given a friendly and slightly zany tone of voice and hundreds of lines were written. The user journey is also filled with sound effects and even a few Easter eggs (try to ask Archie to tell you a joke, for example). The experience is available on any platform compatible with the Google Assistant, may it be a smartphone, a smart speaker, a smart display...

Given that the project is just being released, we'll be able to observe the first results in the upcoming weeks. We’ll look at the following KPIs:

Quantitative :

- DAU / MAU

- Average time spent on the Action

- Return rate (i.e. percentage of 2nd timers)

- Abort rate

- User acquisition growth

- Action ratings (average, over time)

Qualitative :

- WOM (social media)

- PR

Our benchmark will be the leading news Actions in France and around the world.