Building a cross-section article network with AI
December 15, 2021
SCMP Insights
You arrive at a South China Morning Post article, begin reading, then finish it. What’s next?
From our perspective, we hope a reader continues on to another article, not just because it helps our engagement metrics, but also because our mission is to lead the global conversation about China, and that conversation doesn’t begin and end in just one article.
Reading more from across the SCMP gives context and background, filling in the gaps as we attempt to tell the China story to our worldwide audience.
So when we were invited to take part in the 2021 JournalismAI Collab Challenges, a global initiative that brings together media organisations to explore innovative solutions to improve journalism via the use of AI technologies, we saw an opportunity.
With the support of Bennett University’s Times School of Media and Computer Science Engineering Department, the Apac cohort of the Collab Challenges focused on this problem statement: “How might we use AI and audience insights to help newsrooms design more relevant and interactive news narratives?”
The cross-departmental SCMP team - deputy digital editor Shea Driscoll and data scientist Adrian Lam - was drawn to “news narratives”.
Could AI help us choose better stories for readers to read next? We want to give them broader news narratives, offering more insight into the countries, issues or conflicts they read about.
The more diverse the content, the more likely readers would subscribe.
But what does “better” mean in this context? We were inspired by a Google News Initiative Latin America report that found that the more topics and sections a reader read across, the more likely they were to subscribe to your publication. Essentially, the more broadly they read, the more likely they would be loyal readers.
Our story recommendations are typically done with stories in the same section. You’ve read an Economy article, now read another one. It’s often the default selection as an article is built in our publishing systems, largely because we don’t have the tools to help us make more nuanced selections.
So we built one.
The tool we built quickly to recommend cross-section articles.
Working with Initium Media, we developed an AI tool to help with content recommendations in the newsroom. It was created in Google Sheets to get it off the ground quickly. The tool does two things:
Uses collaborative filtering to understand which sections are most closely related to each other
Recommends trending articles based on story performance
The SCMP’s editorial digital team used the tool to make story recommendations on certain articles over four weeks in October and November 2021.
Readers saw the next-story recommendations in two locations in the story. The first was midway through the first article, on the right on desktop.
A “read more” article is placed midway through the first article, on the right.
The second was the next article in the infinite scroll mechanism. When a reader reaches the end of one story, they are immediately presented with the next article, and that can be pre-set by production editors.
The same story is featured in the infinite scroll function.
The results were outstanding. We saw uplift across almost all metrics, comparing readers of more than one article who were exposed to our cross-section recommendations against those who were not.
Readers who were shown more sections and topics visited the SCMP more, on more days, and read more stories per visit. They had a 75 per cent higher likelihood of returning to SCMP.com, while the likelihood of them exploring a third section of our site more than doubled.
It’s very promising stuff, and we’re already aiming to make improvements.
One is to programmatically embed alternate sections’ articles in our automated content recommendation algorithm. Another is to modify our content recommendations based on a user’s visit frequency - we could show them content within the same section on their first visit, then start presenting articles from other sections after their second visit to increase stickiness.
All in all, it has been a very rewarding journey. Figuring out an angle to attack the problem statement that fit with our business needs was a tricky process, but once we zoomed in on content recommendations, everything fell into place.
Crucially, we now have a tool that can be rolled out across the company, benefitting both readers and editors. And most excitingly, the work done here feels like just another preliminary step - the intersection of journalism and AI is pregnant with potential.
JournalismAI is a project of Polis – the journalism think-tank at the London School of Economics and Political Science – and it’s sponsored by the Google News Initiative. If you want to know more about the Collab Challenges and other JournalismAI activities, sign up for the newsletter or get in touch with the team via [email protected]
Article co-authored by
From our perspective, we hope a reader continues on to another article, not just because it helps our engagement metrics, but also because our mission is to lead the global conversation about China, and that conversation doesn’t begin and end in just one article.
Reading more from across the SCMP gives context and background, filling in the gaps as we attempt to tell the China story to our worldwide audience.
So when we were invited to take part in the 2021 JournalismAI Collab Challenges, a global initiative that brings together media organisations to explore innovative solutions to improve journalism via the use of AI technologies, we saw an opportunity.
With the support of Bennett University’s Times School of Media and Computer Science Engineering Department, the Apac cohort of the Collab Challenges focused on this problem statement: “How might we use AI and audience insights to help newsrooms design more relevant and interactive news narratives?”
The cross-departmental SCMP team - deputy digital editor Shea Driscoll and data scientist Adrian Lam - was drawn to “news narratives”.
Could AI help us choose better stories for readers to read next? We want to give them broader news narratives, offering more insight into the countries, issues or conflicts they read about.
The more diverse the content, the more likely readers would subscribe.
But what does “better” mean in this context? We were inspired by a Google News Initiative Latin America report that found that the more topics and sections a reader read across, the more likely they were to subscribe to your publication. Essentially, the more broadly they read, the more likely they would be loyal readers.
Our story recommendations are typically done with stories in the same section. You’ve read an Economy article, now read another one. It’s often the default selection as an article is built in our publishing systems, largely because we don’t have the tools to help us make more nuanced selections.
So we built one.
The tool we built quickly to recommend cross-section articles.
Working with Initium Media, we developed an AI tool to help with content recommendations in the newsroom. It was created in Google Sheets to get it off the ground quickly. The tool does two things:
Uses collaborative filtering to understand which sections are most closely related to each other
Recommends trending articles based on story performance
The SCMP’s editorial digital team used the tool to make story recommendations on certain articles over four weeks in October and November 2021.
Readers saw the next-story recommendations in two locations in the story. The first was midway through the first article, on the right on desktop.
A “read more” article is placed midway through the first article, on the right.
The second was the next article in the infinite scroll mechanism. When a reader reaches the end of one story, they are immediately presented with the next article, and that can be pre-set by production editors.
The same story is featured in the infinite scroll function.
The results were outstanding. We saw uplift across almost all metrics, comparing readers of more than one article who were exposed to our cross-section recommendations against those who were not.
Readers who were shown more sections and topics visited the SCMP more, on more days, and read more stories per visit. They had a 75 per cent higher likelihood of returning to SCMP.com, while the likelihood of them exploring a third section of our site more than doubled.
It’s very promising stuff, and we’re already aiming to make improvements.
One is to programmatically embed alternate sections’ articles in our automated content recommendation algorithm. Another is to modify our content recommendations based on a user’s visit frequency - we could show them content within the same section on their first visit, then start presenting articles from other sections after their second visit to increase stickiness.
All in all, it has been a very rewarding journey. Figuring out an angle to attack the problem statement that fit with our business needs was a tricky process, but once we zoomed in on content recommendations, everything fell into place.
Crucially, we now have a tool that can be rolled out across the company, benefitting both readers and editors. And most excitingly, the work done here feels like just another preliminary step - the intersection of journalism and AI is pregnant with potential.
JournalismAI is a project of Polis – the journalism think-tank at the London School of Economics and Political Science – and it’s sponsored by the Google News Initiative. If you want to know more about the Collab Challenges and other JournalismAI activities, sign up for the newsletter or get in touch with the team via [email protected]
Article co-authored by
 |
Shea oversees SCMP's editorial content distribution and optimisation, encompassing mobile notifications, homepages, social media, and anywhere our readers might be. Shea Driscoll |
_20211210144428.png) |
With a background in Finance, Operations and Product Management from United Airlines, Adrian brings a wealth of data-driven problem-solving experience to SCMP. Since Jan 2021, Adrian has developed a wide range of data models, ranging from Evergreen scoring to Content Recommendation, providing the organisation with valuable readership as well as content insights. Adrian Lam |
