Using GPT-2 to Generate Absurd Names

Sunday night I was sitting at my computer,  tears welling up in my eyes under the glow of the monitor. Now I know what you are thinking – how is this different from any other Monday Eve? Well this time I was holding back laughter from discovering the amazing Cat Fanciers’ Association and their Top Cats and Kittens. Take some time to look through those, it’s an absolute goldmine of cute cats and insane names. Specifically, there are 1,002 awards that have been given from 2009 – 2019 with only a handful of repeat winners.

Why do I know exactly how many winners there have been? I thought it would make an entertaining dataset for my Computer Science class and got to work converting it to a text file. I figured there could be something useful that could be done with a list of names like “Lushell Hollywoodsunset of a Kitkat” and “Topeng I’vegotnews4you of Purrmatix”. I could never be creative enough to come up with names like that! But an AI model probably can.

The secret behind any solid AI model is a large set of data. The 1,002 names are not that many, but I figure they should be enough, especially since people have built programs to make Pokemon names with less. I do not actually know that much about programming, but I am pretty crafty at Googling so I ended up finding Janelle Shane’s page where she did something similar! 

I wasn’t looking to generate all new names though, I wanted them to emulate the CFA’s style. Janelle Shane did something like that as well! Specifically, she used OpenAI’s GPT-2 model so that not only was she using a dataset of cat names, she was using data pulled from across the internet on general patterns in the English language. Reading through her work eventually led me to Mac Woolf’s Python package to use GPT-2 in Colab and got to work training it on the CFA cat names.

I started with the lowest settings I could so that I could run the program quickly and make sure it works, using the 124M model and lowering the learning rate to 1e-5. Running the fine tuning program already led to some great sample results including:

• JUN-JUICE EYE RASPBERY CAT
• CYBERJELLICE EX LIL LO
• JOY TOWN! CUTE SELFIE
• JUN-JUICE IN THE HEART OF HUGUAR JOO-HAUN
• MISS-MEEM-MEEM DUM-DAMAM
• KANSAS CITY MACHINE GUNDAM
• JOKIPEDES BABY PICKENS
• JACOB HILLS TASTE BLUE BLUE TIL I FINISHED
• KANGY-FUH!
• BENJAMIN!
• RUMMAR OF DOLLY-MADE-OUT-OF-PANTS
• TIPPED BILLS THAT ARISE OF SILVER SCANDAL
• KID SHOWMES 2012 BRONCO BALLEROK OF NU MOON

These were entertaining, but they wouldn’t quite fit in with the actual names on the list. After 1,000 steps of fine tuning, here are some of the names the model could come up with:

• JUBILEUM BOWL OF WILLIAM LAFITH
• CHAMY’S CALL ME, BLUE MISTER
• CHANCERY’S QUIERE HAN SHI MAN OF CAT
• FEMALE CARICATURE ENTHUSIAST
• SUNNY WHISKEY BRIGHT RALTAR
• TICKETYBOOK ULTIMA LAND EX
• MARCUS LILY BEAT OF TOMMY LOUIS
• CATBERY TAIL WHIMSICAL AS CAESAR
• SUMO POSELLIUS INFINITE DREAM?
• CHAMSEY’S TUFF TIGER

Much less entertaining, much less chaotic, but much more in line with the actual names! In fact, I would argue if I mixed up the names generated by the trained GPT-2 here with actual ones, you wouldn’t be able to tell the difference. Of the following names, “SUNCOON FAST EDDIE OF FUJICATS”, “PERSIPALS A KIND-OF-MAGIC”, “HOOBEE HELIOS OF DOLL VILLA”, “NAKELI EMMA OF WISHES” and “BENHANA HAN PENG”, only two are real but can you tell which? (It’s the first two.) I saved the results from my model here for your entertainment.