How Devex used MonkeyLearn to scale it's content curation process

About Devex

Devex is a Media platform dedicated to the development community. With a mission to help the aid and development industry do more good for more people, they are also the largest provider of recruiting and business development services for global development.

The story

Part of Devex operations involves publishing great content for their vast community of avid readers. This content can be found on the home page or in their news dedicated section.

Being already customers of MonkeyLearn for processing documents and reports, they knew the power of Machine Learning for text analysis and decided to reach out with a new challenge.

The Challenge

The process of curating content involved manually reviewing their different feeds (consumed in Slack through Feedly) to identify which content was deemed relevant for their community.

• On a given week Devex could get around 3,000 different pieces of content, out which only 300 would be tagged as worthy.
  
  

• 
  This process took around 10 hours each week to complete.

The Solution

• Using Zapier, Devex gathered a list of all the RSS entries throughout a week, about 3,100 entries in total. With a different zap, they gathered a separate list of entries that were manually classified as relevant, about 320 in total.

With this data, Devex was able to build a MonkeyLearn custom classifier that would tag the content for them as either relevant or not.

Then using MonkeyLearn's Google Sheets add-on, Devex processed the data and then compared the results to a test set manual classified.

The initial version of the classifier correctly classified 75% of the entries, right out of the box, without any retraining. Of the incorrectly classified 25%, 20% were false positives, and 5% were false negatives.

After a few iterations to further improve the accuracy of the model, Devex connected their Feedly feed through Zapier to MonkeyLearn and then to Slack so that they could now manually read only the entries the classifier deems relevant, or around 33% of all entries, instead of reading 100%, saving precious time that can be spent on more important tasks.