Case Study

Context Window: You can't spell ‘metadata’ without EDM

By Sonali Verma

November 4, 2026

Hello everyone,

Have you ever found yourself in a situation where you are about to go on stage to speak at a news media conference – only to discover that you are scheduled immediately after a DJ who was spinning dance tunes and had everybody on their feet, grooving to the beat? 

That’s a tough act to follow. Oh, and the topic of your panel is…. taxonomy. Er, ok, what could be more fun than a panel on taxonomy? (If you said ontology, hats off to you!)

Yet the Lenfest AI Fellows pulled it off with aplomb, inspiring some great questions from a highly engaged audience. As one of our Fellows, Peter Rasmussen, remarked: “It’s the perfect set up. You can’t spell ‘metadata’ without EDM.”

Sign up for Context Window

Context Window is The Lenfest Institute’s newsletter sharing insights and lessons from the AI Collaborative and Fellowship.

Subscribe

* indicates required

Email Address *

Getting classification right at The Baltimore Banner 

Speaking at the News Product Alliance Summit, Peter described The Baltimore Banner’s first major AI project: The local news startup wanted to understand who was reading their stories and what readers needed, but the existing website categories came from newspaper sections (such as sports or politics) that mostly described where an article was placed, rather than what readers sought from it.

“To go deeper and to elicit user needs, we needed a taxonomy, a set of classes that got at user needs,” he said.

They used taxonomies previously published by the BBC and Financial Times as a starting point and refined them further. 

“If you don’t figure out your taxonomy correctly, everything you classify is going to have to be reclassified,” Peter pointed out.

The process

• Developed the taxonomy iteratively with editorial colleagues. Rather than assuming the initial taxonomy would be right, Peter said the team built its own through an iterative process. They started small and expanded as confidence grew, limiting unnecessary reclassification and expense.
• Evaluated classifications with editors. He described testing a prompt-based classifier on a sample, having editors assess the labels, and keeping a held-out “gold test set” to measure performance more objectively.
• Used the approach for practical newsroom needs. Peter described building a news monitoring tool called “News Detector” to flag events for editors on sites that they otherwise had to check individually. He also used a simple LLM classifier to filter obvious spam from incoming tips.

Challenges along the way

• Delays: The development process took much longer than Peter expected because of the surrounding coding setup and complexity.
• Cost: He warned that classifying stories can add up quickly, which was another reason to begin with a small amount of data.
• Ongoing maintenance: He emphasized that a robust pipeline is not finished when it first works; it needs monitoring, alerts and upkeep. It is a product that must be maintained. 

Collaboration with stakeholders

Peter said he works closely with colleagues in editorial and analytics and shares findings with them quickly, often through Slack.

Editors also contribute to defining and evaluating the labels. Peter described their existing tags as valuable labeled data that could serve as a standard for assessing whether a classifier reproduces editorial judgment.

Taxonomy frees up advertising inventory at The Boston Globe 

The Boston Globe tackled taxonomy from a different angle, their AI Fellow, Omar Reid, told the audience. 

He described a brand-safety problem: Advertisers wanted their ads kept away from unsuitable stories, but the Globe’s existing section-level categorization was too broad. After an advertiser’s ad appeared beside a tragic story, the team realized it could not reliably identify and avoid similar contexts across all its articles, so it blocked whole sections.

“The advertiser doesn’t care about the section level,” Omar said. “For that, we need a granular understanding of our articles.” 

For example, a section such as Metro could contain very different stories. “You can have ‘this restaurant has opened,’ or you can have ‘someone got stabbed on a bus,’” he pointed out.

A more nuanced taxonomy has meant ads can appear against a wider range of articles, unlocking revenue for the Globe.

The process

Omar’s approach was to centralize the Globe’s content, then classify it and use the resulting metadata in products. He used the IAB’s content taxonomy.

With content centralized in Snowflake, he built three applications for three purposes:

• Brand safety: Newly published articles go through a pipeline that analyzes them so the team can determine whether they are suitable for particular advertisers.
• Journalist search: A Slack bot lets journalists search Globe articles from the centralized data store, addressing the frustration of having to use Google to find their own work.
• Personalization: By classifying past articles and examining reading histories, the Globe could develop personalized newsletters.

Challenges along the way

• LLM confidence scores weren’t dependable. Omar cautioned that asking a language model to predict its own confidence did not produce useful, calibrated scores. He said they needed other kinds of models for more “sensible” confidence estimates. 
• Classification required labeled data and iteration. He said the team had to build a clean, “gold standard” labeled dataset, evaluate model changes with metrics, and keep iterating. This work was unfamiliar to some colleagues accustomed to more predictable software engineering. 
• Building buy-in across teams: Different engineering groups had their own roadmaps, so applying classifications across publications meant coordinating with those teams, rather than assuming they could simply adopt the work.

Collaboration with stakeholders 

Omar said he began by speaking with people across the business to understand what they actually needed:

“When I first joined, it was my goal to talk to as many people from all sides of the business, so that I could really understand what they need,” he said. He ended up working closely with editors and development teams, explaining the value of the project and fitting the work into their priorities.

He emphasized that successful collaboration meant understanding organizational relationships and winning support through persuasion rather than relying on authority:

“It’s better to use soft power,” he said.

Tools you can use

• IAB Content Taxonomy Overview (IAB website)
• IAB 3.1 Content Taxonomy (Github)
• Omar’s data map for the newsroom to help them understand which data they have access to and what they could search for. 
• Peter’s set of taxonomies at The Baltimore Banner.

Coming up in the next edition of Context Window:

How Chicago Public Media used metadata across its archives — and built a brand new product that is attracting rave reviews. 

Have any questions on how to build similar tools for your own newsroom? Please don’t hesitate to reach out to [email protected]. If a colleague forwarded this email to you, you can subscribe to Context Window here.