in Projects, Programming, Technology

Generating startup names with a neural network

Naming a startup is hard. We spend a lot of time on it at Pioneer Square Labs. I was inspired by Andrej Karpathy’s attempt to generate baby names with a Recurrent Neural Network. What would happen if we applied that approach to startup names? The results were hilarious.

Here are some of the names I generated. This was a small run of 2,000 generated words filtered to 614 available domain names, ranked by their Levenshtein distance (a simple character distance) to dictionary words, and by scrabble score. From those 2,000 – here are my top 5 (available at the time of writing).

  1. BeepGram.com
  2. EverHint.com
  3. Boomata.com
  4. CinnePix.com
  5. CarJus.com

There is a lot of noise, but Sireze (that’s the generated name I chose for my name-brain) can run all day generating names. For every 1,000,000 of them, there are about 2,500 interesting names. Basically, the RNN steps from letter to letter, picking the next letter based on the frequency of terms it was trained on. It’s like a baby babbling things that sound like what it has heard before. I think that’s pretty great, but this technique needs refinement if it is going to only suggest good names without the help of a human filter. I haven’t taken the time to learn how to change the RNN parameters to give me better results.

Here are the instructions I followed on how to install and train Torch-rnn, a set of recurrent neural network tools for single-letter based learning and input. I trained the network on lists of startup names exported from Crunchbase, and quite quickly had a baby babbling startup names.

Next, Greg Gottesman encouraged me to score the names on more dimensions, for example:

  • Is the dotcom available?
  • What’s the scrabble score? Rich Barton has some great tips on what makes for a good brand nameHigher point letters (Z, Q, X, J, and Ks) are more memorable. The fewer syllables the better. Add points for alliteration or double characters
  • Does it contain a latin root (or, how similar is it each syllable to an english word)?
  • Do the syllables that rhyme?

While Sireze doesn’t generate names that have any special insight behind them, it can generate lots of names that are inspiration for your own creativity (and they are certainly fun to read). Of course this approach would work well beyond startup names. How about products like pharmaceuticals, cars, and rap stars?

Will you buy any of the domains on this list? If you’ve got something you’d like me to name, send me a list of 1,000 similar names and $100. I’ll run it through Sireze and see what she comes up with.

 

Leave a Reply for Murat Cancel Reply

Write a Comment

Comment

  1. Cool. I think another valuable dimension would be whether it sounds the way it is spelled. Names with superfluous letters or weird spellings aren’t as good as those without.