sauravkaushik8/startup-name-generator

Script to randomly generate company name ideas

startup-name-generator

Naming a startup is hard.

I therefore wrote a Python script to randomly generate company name ideas.

It takes an arbitrary text as input, and then trains a recurrent neural network (RNN) on each its words, learning the structure of the text. The input text can be a simple word list (e.g. Greek or Gallic), or a chapter from a book, or just a random list of words (e.g. all Pokemon). The script then generates new random names that sound simliar to the provided list.

I provided a few input texts in the wordlists directory - see the README.md there for descriptions. I also kept the initial Jupyter Notebook here that I used in the first development stages. It walks you through the inner workings, if you want to experiment yourself.

You can (and maybe should) save a trained model in the models subdirectory, so that you don't have to train it every time you want to generate new names.

Detailed description and tutorial

I wrote a blog post with a more detailed description of this project, and a few sample calls and outputs.

Usage

• Show a brief help:

./generate.py -h

• Sample usage, appending a custom, fixed suffix to each name:

./generate.py -e 500 -n 10 -t 0.7 -s models/gallic_500epochs.h5 wordlists/gallic.txt --suffix Software

Toartio Software
Riuri Software
Palia Software
Critim Software
Neibo Software
Veduos Software
Eigla Software
Isbanos Software
Edorio Software
Emmos Software

(I found a long list of possible suffixes here)

• After you stored the model (with the -s option), word generation is quicker when you load instead of re-compute the model:

./generate.py -n 10 -t 0.7 -m models/behemoth_500epochs.h5 wordlists/behemoth.txt --suffix Labs

Literature/References

• Andrew Ng's Deep Learning MOOC
• http://karpathy.github.io/2015/05/21/rnn-effectiveness/
• https://github.com/keras-team/keras/blob/master/examples/lstm_text_generation.py

TODOs / next steps

• When using a stored model for simulation, I still need to load the wordlist in order to generate the character dictionary. The wordlist can not be changed, otherwise you'd have to retrain the model. It would be nice to store the character set (it's the ix_to_char dictionary in the code) along with the model.
• Since I'm not yet a Python expert, there are most likely some suboptimal ways of doing things in the code.
• I currently filter out the hyphen during preprocessing. Ideally, I should keep it if it appears within a word, and filter it if it represents something else like a bullet list item.
• It would be cool to have an option to specify that one input name should be one line instead of one word. Then, one could use lists of actual company names that include symbols like ampersands, whitespace, etc., and sample these as well. See here and here for nice possible input lists.