Context Cards: Amazon Prime X-Ray, but for News
The post was originally published at JournalismAI Blog on the London School of Economics website.
Context Cards is a machine learning solution that creates and suggests context — data, bio, summary, location, timeline — as short-form content in long-lasting news cycles to audiences and journalists, alongside an article. The model trains on newsroom archives, and learns from editors’ feedback.
The closest similar product is…
When you watch a movie on Amazon Prime Video, there is a button that comes at the top of the video called X-Ray. On clicking, it provides context to the audience:
- In scene, cast, characters: X-Ray shows all people involved in a particular scene, the entire cast and bios of specific characters.
- Scenes: They also break down the entire movie into a timeline of scenes, allowing users to jump to specific parts.
- Trivia: Interesting details about the movie.
Why short-form content
Coming into the JournalismAI Fellowship, we already knew that audiences:
- Aren’t reading long-form fully
- Prefer to read short-form content that can be digested quickly
- Aren’t necessarily aware of the context behind long-lasting topics
Hence, TOI spent the better part of a year building NewscardCMS — a platform for developers to create modular content templates and for editors to author them. You can read more about Newscards on Medium.
Why automate with AI
From deploying NewscardCMS, we learned that:
- It isn’t straightforward for editors trained in writing in the inverted pyramid style to author content in cards
- The newsroom workflow chases what’s new. Hence, the workflow doesn’t allow for updating evergreen cards
- It also isn’t obvious for the desk to attach existing evergreen cards to stories
Hence, it is a worthy goal to use AI to automate not only the creation but also the plugging of context cards.
We decided to focus on long-lasting and slow developing topics instead of fast-developing and possibly short-lived topics.
Editorially speaking, many of the development and divisive political issues we face in our information space are long-lasting, slow-developing topics. For example, stories related to gender rights in Taliban ruled Afghanistan will last for years to come.
It also meant that our algorithms wouldn’t be live pushing content to audiences. Our editors would have the time to supervise and curate the output from the algorithms.
From a business perspective too, we would be spending previous computational infrastructure on topics that have a long arc and thus long shelf life.
Finally, the algorithms that mine nuance and context out of an archive of stories are different from the algorithms that predict early signals and momentum.
Why newscycles and not topics
Generally, a topic in computer science parlance is often related to taxonomy. A taxonomy has a fixed set of buckets in which content can be classified. To experience taxonomy or topics, browse through your Twitter feed and you’ll find tags like the one below highlighted in pink.
However, what we are interested in is news cycles. To experience a news cycle, let’s look at Twitter again. On the top right side corner, Twitter showcases the most trending news topic. For example, protests in Iran.
And when you click on it, it takes you to an aggregation page dedicated to that news.
Building the case for our stakeholders
To avoid disturbing the existing product, we intend to produce context as small widgets that can go on existing pages or create completely new pages. For example, the timeline could look like this:
The project also needed to serve clear product and business goals. Hence, we decided that all output from the project should serve one of two metrics: Pages Per Session and Sessions Per User.
- Sessions Per User: Features that get audiences back to the platform.
- Pages Per Session: Once they are on the platform, get audiences to consume more content
If the feed had a follow button, then the Verge could have sent out a notification or email to those who follow this topic to increase sessions per user.
Progress so far
We’ve split into two teams:
We hired a consultant —Anuj Karn — to build out a Named Entity Recognition algorithm.
Meanwhile, our lead data scientist and project partner, Karn Bhushan, started exploring topic modelling algorithms with guidance from Dr. Tess Jeffers, Director of Data Science of the Wall Street Journal. We were able to bring the accuracy to around 70%.
The biggest challenge that we foresee is translating the output from topic modelling algorithms (topics) into news cycles. Let me elaborate:
From topic to news cycle. Topic modelling are unsupervised algorithms that solve for similarity i.e. they find clusters (buckets) of similar content.
- The output from topic modelling will need to be given to editors to derive meaning from it.
- Editors interpret if one or more topics from the algorithm add up to an news cycle.
- The news cycle then needs to be labelled (headlined).
Refining the news cycle. We will need to give editors the ability to refine the output.
- False Positives: The algorithm decides that a story is part of the topic but it really isn’t.
- False Negative: The algorithm decides that a story is not part of the topic but it really is.
Maintain backward compatibility. To find new topics and thus news cycles, it becomes critical we retrain the model with new stories and the False Negative and False Positive tagging.
- However, when we retrain the model spits not a completely new set of topics.
- Hence, we need the ability to find nearness between the new topics and the old topics so that all the manual tagging and labelling can be carried forward.
- Created by Krista Quiroga from Noun Project
- Created by H Alberto Gongora from Noun Project