Pandata Blog

Approachable, Ethical, Human-Centered AI

Back by popular demand – our HolidAI Greeting Card Generator has a fresh look, along with updated images and messages for the 2019 holiday season! And to be honest, it’s gotten even weirder.

Machine learning and AI can be confusing and scary subjects.  We want to make them fun and accessible so we decided to use artificial intelligence to generate holiday greetings that are merry and, well, memorable.  So how did we create it?

The majority of data in today’s enterprise lives as unstructured text. Natural Language Processing (NLP) is a rapidly evolving discipline that transforms unstructured text into meaningful information. We’re all familiar with autocorrect and autocomplete that has made cell phone messaging a more user friendly experience. Customer experience teams are leveraging advances in NLP to more effectively react to the voice of the customer. With customers expecting quicker response times and reactive customer service teams, chatbots are becoming increasingly widespread. Our holiday greeting generator is an example of the technology enabling chatbots to learn from examples and respond to customer requests.

To arrive at the AI generated holiday greetings, we used a Neural Network that “learns” patterns in language and then generates new messages. This kind of neural network is called a textgenRNN and it generates new words one letter at a time. We started by scraping around 4,000 Christmas, Hanukkah and holiday card messages from greeting card websites. Then, we “taught” (or trained) the neural net on these messages. After each round of training, the model compares its output with the training data set and then modifies its process to make it more like the training data. For more on Neural Networks, check out this blog from earlier this year.

The generator learns everything from capitalization to punctuation and spacing, sometimes even making up its own words. Much like a baby learning language, the algorithm first generated nonsense. After one round of training, an example card message is:

“Hoorre than pronace.”

After a few more rounds of training, the output started to make a bit more sense, at least generating recognizable words, mostly. Two examples:

“Happy Hanukkah
and a Hanukkah
and a Happiness this Christmas sime to shine the season and happiness and happiness and all the time to season and happiness and happiness and a Hanukkah
and a Hanukkah season the season and happiness”

“ParaChristmas Duck all for reason, paro little a family merrie family gifts.”

After around 6-8 rounds of training, familiar messages began to appear:

“Wishing you a holiday filled with peace, joy and happiness.”

Along with some more nonsense:

“Wishing A count perfect a hall!”

After 15 rounds of training, the majority of the messages seemed to be reasonably coherent and on topic. Much more training and the neural net started to memorize the cards it originally was trained on and just spit them back out, which is not as fun as the cards you’ll generate here.

As the traditional AI generated holiday greeting goes, “wishing you means lights for the new year out of your minions!”

Hannah Arnson is Lead Data Scientist and Nicole Ponstingle is COO at Pandata.