Deepnews in COVID-oriented mode

Original article was published on Deep Learning on Medium

Deepnews is about picking out the quality news from the web. For a long time, the conversation will be about the COVID-19 pandemic and its aftermath. We took several initiatives to address that.

Our goal at Deepnews is to spotlight the best articles from the web by using machine learning algorithms. The main application for it is a set of 11 newsletters covering the following issues:

Here is what is new:

To make our selections more accessible and in sync with a news cycle overwhelmed by the COVID-19, we took a few steps:

• We redesigned and simplified our website which you can look at here:

  • We made our product much easier to test by removing the need to enter your credit card. You can now evaluate our products for a full month without fearing unwanted charges at the end of the period.

Better access to our editorial products is a response to the huge demand for information about the crisis. It is consistent with the strategy of large media outlets that unlocked their paywalls for COVID-related content. All of them saw a record increase in new subscriptions, as shown in this chart by Piano.io:

It is unclear how many subscribers will remain at the end of trial periods, but the hunger for reliable information is higher than ever and will last for a while.

More COVID-centric editorial in Deepnews newsletters

Our free Deepnews Digest is now bi-weekly and entirely devoted to the crisis, each week offering a specific angle to delve deeper into subjects often treated in passing:

The principle is the same as before: we scan stories from more than a thousand vetted sources, our Deepnews Scoring Model scores them, and we serve a selection of the 25 best articles. For all our newsletters our editor-in-chief Christopher Brennan and our second editor Shrey Bhandari eliminate unavoidable mistakes of the deep learning model (false positives or stories completely off-topic), they add a few “editor’s notes” and they hit the send button.

From an editorial standpoint, our newsletters are reflecting the overwhelming trend in the news: many of the articles we select are somewhat related to the pandemic.

They also respond to another request from the audience: the need for expertise. As expected, the pandemic has triggered an explosion in misinformation. It ranged from blatant fake news propagated on social networks to more pernicious forms of punditry that either fed panic or entertained false hopes.

Here are some that have been contaminated (pun intended) by COVID related news:

• “Matter of Facts,” our product dedicated to misinformation, covers extensively the mechanisms in play, from antivaxx groups to governments, and the efforts deployed to debunk them.

• “Gig Economy” covers a sector terribly impacted by the economic downturn and explores the sub-segments that are booming such as food delivery.

• “Future of Medicine” is dominated by COVID-related stories, ranging from supply chain issues to contact-tracing, or the race for a vaccine.

• You might even find some stories related to the impact of the crisis in “Women in Tech” (one of our most talked-about products) and in “Green Energy”. Both cover critical issues that might be impacted negatively by the fallout of the pandemic.

We are also continuously adding sources. As of yesterday, we had 1900 English-speaking publications, with more to be evaluated in the coming days as we prepare new newsletters.

Source additions come from those sites referenced on a regular basis by websites that we already monitor and trust. We also look at syndicated sources and we set up our own automated monitoring. Finally, we take into account readers’ proposals especially for niche/expert publications — a segment we want to develop reach out to us if you have any proposals to make).

Working on a French version

We started Deepnews in English because of its origin (during my time at Stanford’s JSK Fellowships) and also for market size reasons. But we always knew that we would jump at the first opportunity to bring the concept into other languages.

This spring, we received a grant that will help us to develop a French version of the platform. The adaptation of the main algorithm, the Deepnews Scoring Model (DSM) will start in the coming weeks.

We will have to deal with a nagging question: English-speaking journalism is extremely structured with narrative codes that are well-defined. Scores of books about news reporting and stylebooks published by news organizations and journalism professors explore the subject in great detail. In France (and I’m afraid in many Latin countries), journalism is less disciplined in its construction.

The notable exception is Agence France-Presse, which is keen to maintain a well-structured flow of articles. It is a matter of principle but also dictated by the fact that AFP’s output is available in six languages.

For the most part, however, in French, writing skills — in the literature-talent sense — will often take precedence over clarity. To put it another way, training a large neural network and applying natural language processing (NLP) techniques might be difficult for French journalism. That indeed will be an interesting challenge for our data science team.

A group of French publishers has agreed to join the project by providing large volumes of articles to train multiple versions of the DSM that will be required to determine the right formula (more on this in a few weeks).

Unchanged roadmap: surfacing great journalism on a large scale with multiple dedicated and automated editorial products

Despite economic uncertainties, our roadmap remains roughly the same. The only project we shelved, for the time being, is our ambitious enterprise platform aimed at the global information sector. It was supposed to allow information producers and distributors — not only media but all corporations that produce and curate large volumes of articles — to access our algorithm in a self-service mode, to score articles. In doing so, they would be free to build whatever they want: recommendation engines, advertising, digital subscription tools, and automated curation systems, etc. This project is capital-intensive while the demand from potential customers might stall for at least a year. Hence the postponing of the development phase. Nevertheless, as our technical needs increase, we hired a CTO, Girish Gupta, a physicist and computer scientist by training, who is also an experienced investigative journalist with Reuters.

Deepnews has now spread over 10 time-zones: Bordeaux (where our CEO, David Finch, is based), Paris (the data science team, me), Brussels (Christopher Brennan, editorial), Mumbai (Girish, the CTO) and Auckland (Jessica Burkhead, our recruit from last week, will be in charge of marketing). Hence my interest in remotely operated companies (see last week’s Monday Note featuring a 185-person company operating in 40 countries).

At the beginning of the summer, we will be launching a new batch of paid-for newsletters. Ideas on the topics we will cover remain sketchy at this time. But again, how the post-COVID era — or, rather a COVID-integrated existence — will unfold will be critical in our choices. We can assume broad changes in many areas like work, travel, housing, urban life, the macro-economy, management of inequalities, etc. We might ask for your opinions, dear Monday Note readers, in a few weeks. Until then, be well.

frederic.filloux@mondaynote.com