By Chris Testa
Director, Engineering
Adly, Inc.

Earlier this month, it was my privilege to present “How Hollywood Learned to Love the Semantic Web” at SemTech 2011, the conference on semantic technology.

Here’s the story behind that presentation and today’s video interview and “How to” guide on SemanticWeb.com.

In late 2010, we at Adly had hit a scalability wall.

We already had 1,000 celebrities in our Twitter endorsement network, and I anticipated there’d be thousands more to come as Twitter went mainstream around the world.

We needed to automate the process of matching those 1,000 celebrities with our 150 brand advertisers — a process that had previously been highly subjective, tapping our team’s tribal knowledge.

Further, our system needed to be designed to scale to accommodate the new celebrities we were attracting every day.

I needed to find — and gain programmatic access to — in-depth profiles and brand-affiliation data on thousands of celebrity artists, athletes and experts.

Enter Linked Data

I had done a project at University of Maryland with Prof. Jim Hendler’s MINDSWAP program during undergrad. But I was skeptical of most Semantic Web technologies as being frankly overwrought and too complex for my everyday use.

Through both my IBM Extreme Blue internship and work at Google and YouTube, I’d seen little if any practical application of SemTech on the job.

But with a pressing business need and few resources, I started exploring the freely and open Linked Data sets that might help us easily aggregate and integrate brand affinity data.

A Treasure Trove

In a review of the Linking Open Data Cloud, I identified Freebase’s celebrity pages — complete with “wears,” “shops at,” “eats at” data and more — as a treasure trove for brand matches.

From musicians like 50 Cent and TV personalities like Audrina Patridge, to athletes like Serena Williams and Nick Swisher of the New York Yankees, Freebase made it easy to access, vet and ingest extensive profile data.

We were quickly able to bring in everything from basic profile data (profession, age, marital status, etc.) to existing brand affinities with clothiers, cars, resorts, retailers and more.

Let the Games Begin

But before I could trust the Freebase data, I needed to accurately match each Freebase profile with the right Twitter identity — a process called reconciling.

It turns out that humans clearly beat computers in one key area, which is saying when things are the same and when they’re different.

So we created a reconciliation matching game pairing Freebase profiles with @Names in a simple “match” vs. “no match” fashion.

We invited our expert talent team to play, sorting the celebrity wheat from the mis-match chaff at least twice for each @Name until the reconciliations were done.

The game made easy work of the initial 1,000 celebrities, and has made it easy to onboard new celebrities as well. (Co-creator Jeff Schenck goes as far as to call it “beautiful and elegant.”)

Best of all, comments from the talent team included, “that was fun” and “when can we do that again?” We are contemplating publishing the game to tap the power of fans.

A Scalable Solution

Within weeks, we had put Freebase to work in a scalable solution that brought profile data, summary bios, professions, etc. directly into our BI Interface, “Blingalytics.”

That made it easy for the talent team to filter and sort celebrities by a number of attributes, including gender, age, parental status (yes / no), etc.

For the first time, the team could do such seemingly simple tasks as quickly sort all of the American Football players in our network into a SuperBowl campaign, and then rank them by performance (ability to drive consumer traffic on the Web).

Without phone calls or emails to agents, the team could find all of the actresses in our network that are moms and reach the decision-making moms on Twitter that CPGs want to reach.

Now six months into the project, we are contributing corrections and updates back into Freebase, and have our eye on creating a robust celebrity affinity graph.

Click here to read more about the process I followed.

Better Brand Matches Make for Happy Advertisers

In business outcomes, we have been able to automate the initial matching of the now 1,200 celebrities in our network with the brands we serve.

Not only did we streamline the process of celebrity selection and make it easier to avoid brand conflicts, but we also enabled the talent team to make surprising or counter-intuitive matches.

For instance, who knew Adam Corolla was a carpenter for 13 years before becoming a radio show talk host? He could endorse brands like Home Depot, Lowes, Ryobi or Stanley, etc. with authority.

Most of all, better brand matching has resulted in more effective campaigns, from superior audience engagement (conversations inspired) to increased campaign performance (retweets, clicks-through, etc.).

Check out my video and 5 tips to get started on SemanticWeb.com and let me know your thoughts on Twitter @crstesa – hope to hear from you soon.