word-maker

makeword.py implements an n-gram language model to generate words that (most likely) don't exist.

Words are generated on a character by character basis. The likelihood of a character being generated is equal to the relative frequency with which that character followed the last n-1 generated characters in the training set. The first n-1 characters are generated in one chunk drawing from a separate collection of (n-1)-grams that appear at the beginning of words. By default, n = 4.

The default training set is an English dictionary of ~450,000 terms. There are separate entries for each word's different possible prefixes and suffixes, such as run and running.

USAGE

$python3 makeWord.py FLAG(S) ROOT

Flags and root are optional. Flags and root will modify the functionality of makeWord.py.

FLAGS

• -c
  • print generation with green and highlights denoting particularly expected or unexpected generations respectively
  • see COLOR CODING section below for details
• -i
  • print the relative frequency and relative 'expectedness' for each character generated
  • see COLOR CODING section below for details
  • implements -v
• -m
  • wait for user to hit ENTER to generate next character
  • implements -v
• -max NUM
  • set maximum length of generated word to NUM
• -min NUM
  • set minimum length of generated word to NUM
• -n NUM
  • set n-gram length to NUM
  • longer n-grams tend to yield more convincing fake words but are more likely to generate real words
• -t NUM
  • wait NUM seconds to generate each character
  • implements -v
• -v
  • print in-progress generation every time a character is generated

ROOT

User can pass a string of length >=n-1 as the final command line argument to serve as the "root" of the generation. makeWord.py will build off of this root to generate a word.

EXAMPLE EXECUTION

In this example, 'rad' is passed as the (optional) root.

$python3 makeword.py -max 12 -v rad
<Hit ENTER to generate word>

rad
radi
radic
radicr
radicro
radicrol
radicroli
radicrolis
radicrolis

COLOR CODING

If the -c flag is passed, characters may be highlighted green or red to denote 'expected' and 'unexpected' generations respectively. A generated character's expectedness is determined by how the probability of its generation compares to the average probability of all other possible genenerated characters given the last n-1 characters. The brightness of the highlight corresponds to the magnitude of its (un)expectedness.

For example, let n=4 and 'xyz' be the last n-1 characters of an in-progress generation. If, in the training set, 'a' follows those characters 90% of the time and 'b' follows them 10% of the time, the average relative frequency in that state is 50%. In this case, 'a' is 40% more likely than average. Thus, if 'a' were generated, it would be shaded a bright green.