What happened when we used machine learning to rename Armadillo?

Nicholas Blake-Steele

How our agency came to be called ‘Armadillo’ is a legend lost in the myths of time. Some in the office say the strength of the impenetrable armour was an inspiration to the founders when we started out way back in 1992. Others believe it was randomly chosen from a dictionary after a few too many drinks.

This got us thinking: which of the technologies that hadn’t been readily available in 1992 could we use in 2017 to help us (theoretically) name Armadillo again?

Machine learning was an obvious fit for us – we’d already built a platform to test machine learning internally to work on subject lines and help us recommend alternative approaches to improving customer engagement with emails.

With this in mind, we collated a list of 1,500 common animal names in the English language and politely asked the machine learning powers to work their magic.

Stage 1: the learning phase

Sample output:

  • Basyhenllhflen morpunk
  • Pown sgoh
  • Caek (regposked bommel)
  • Cooba
  • Araten (cyan

At this stage it’s clear ML is working out the patterns which form an animal name – it knows it needs spaces when it has multiple words, and it’s even begun to work out that some words require brackets (these are often descriptive words to denote slight variations in the same animal classification). However, almost none of the words are recognisable English or even possible to pronounce – ML still has work to do.

 

Stage 2: the improvement phase

Sample output:

  • Horned dianon gebin
  • Pumkla’s blue-fanted barres
  • Common racossis goby
  • Seven-mailled fox
  • Pabralian tortontail skail

It’s a good step forward: ML is now beginning to correctly spell some common English words often used in animal names. However, it still has quite a few words that are very clearly made up with no foundation in the English language, meaning they probably wouldn’t be readily adopted when David Attenborough discovers the next deep sea species!

 

Stage 3: final phase

Sample output:

  • Mawk-easted monster
  • Bee-eater (bracking)
  • Red-tailed heaven
  • Common yellow mandaboo
  • Degwlan fork (giant)
  • Jackil (tommen)
  • Long otter
  • Ring-necked owl
  • Black-balled shork
  • Black-throated three-beater

By the end of the learning process, our machine learning engine has made significant progress. The majority of the words are recognisable, with only a few made up – but almost realistic sounding – names thrown in. Office favourites include ‘Mawk-easted monster’, ‘Common yellow mandaboo’ and ‘Black-throated three-beater’..

For a relatively quick process with no tuning of parameters or weighting of favourable words or pattern, we were pleased with the eventual output. We put our shortlist forward to James our CEO, but for some reason he’s yet to approve any of them…

The above is just a tiny (and fairly simple) snapshot of what we believe will be a transformation of how the CRM industry works with and uses data in the coming months and years – and we’re only just getting started!

If you’d like to know more, do get in touch to find out how Armadillo can help your business use its data to understand and reach your customers better.

Recent posts