Is it possible to predict revenue with Artificial Intelligence (AI)? With the content quality scoring system Deepnews.ai, it may be possible to optimise the revenue potential of already great content.
Frédéric Filloux, chief executive officer of the company, gave insight into the progress of Deepnews to attendees at the Local Reader Revenue Symposium that INMA co-hosted with Mather Economics, as part of Media Innovation Week on Wednesday afternoon.
Giving background on Deepnews.ai, Filloux said the system is not a fake news detector. Rather, it aims to measure the quality of an article or the amount of value added that is put into a news story by an editor. The team has been able to train about 100 versions of the current model, amounting to approximately 1,500 hours, to achieve an acceptable accuracy rate.
The first question Filloux asks, which he says is a fairly relevant one is: “How accurate is our model?”
The Deepnews team measures the model’s accuracy by giving journalism students thousands of stories to score and then testing scores against the algorithm. This has reduced the mean square error from 0.8 to 0.4 over a year Filloux said. The confidence index, which results from a threshold beyond which the model acknowledges its inability to safely return a score, is below the critical threshold of no-confidence in 80% of the cases.
In terms of practical applications, Filloux said the first step is to create a scalable, automated newsletter system. Deepnews Digest Newsletter is free and provides the best 25 stories on any given subject. The second is a paid-for industry verticle for sectors in which people are having a hard time finding the right information.
“The whole idea is to have a newsletter that we can launch in just one month in between the decision and the launch of the newsletter,” Filloux said. “And when we’re going to be making the newsletter, we don’t want a human to be spending more than an hour putting together the newsletter.”
Later, Filloux said Deepnews aims to develop a self-serving API based on the idea that publishers submit batches of articles, get a score, and pay a fee based on volume.
“Who are we targeting? Of course news publisher remain our natural and preferred market because we come from this,” Filloux said. “And we know exactly what the needs are. And the other natural targets are news aggregators. Any company for which quality information has a strategy value.”
The idea behind Deepnews was to match ad price with the quality of the news that appears next to it. The news business is an unique sector where there is no correlation between the cost of producing a story and the actual value for advertising, Filloux said. Publishers can also address issues with content recommendation engines, which he calls the lowest hanging fruit in the industry.
The Holy Grail use would be predicting ad revenue and subscription conversion. With permission, Deepnews has been using the data of Mather clients to assign content scores and correlate that information with conversion and path to conversion. The data reveals patterns for most types of content, but there are some content types that are outliers.
“For reasons we still need to determine, business stories do not behave in the same general ways that other stories seem to behave,” Filloux said. “Nevertheless, we have roughly two or three hits that correlate to the actual score.”
Preliminary conclusions for ad predictions reveals the model is dealing with too many variables to claim a reliable measurement. For subscription prediction, the model is much more reliable.
“It definitely works, and this is super important,” Filloux said.
Moving forward, Filloux told the audience that the next step is to improve the Deepnews model by focusing on speed and reducing the number of parameters in the model. Improving the accuracy rate and confidence index is also a goal. Deepnews welcomes collaboration as it continues this journey.
“This is not the kind of stuff we can do alone, hence the relationship with Mather,” Filloux concluded. “And thank you, Mather, for spending time with us and working with us. We need more data because finding data for deep-scoring for news is very complicated.”