Gem Version

At EXPLO, we often have a need for specific vocabulary-generators. For example, we might want to make a password generator, or a Harry Potter house-generator, or some such thing.

As it turns out, Markov Chains are an excellent way to create specific vocabularies by "training" a model against a set of words to determine common combinations.

While there are quite a few wonderful Ruby libraries that do this, they all focus either on actual English words, or on creating random sentences but not words. We created this library to do the same thing, but with words, hence the name MarkovWords.


Add this line to your application's Gemfile:

gem 'markov_words'

And then execute:

$ bundle

Or install it yourself as:

$ gem install markov_words


Basic usage is as follows:

require 'markov_words'

generator =
# returns a random word
puts generator.word 

You might prefer using a number of n-grams (letter combinations being tracked) higher than the default number (which is 2). We've found that the higher you go, the more accurate words tend to sound, as the likelihood that you've started with a partial word the entire length of a word from your dictionary goes up. The increased "real-sounding-ness" comes at the expense of having to generate a much larger database of n-gram => letter correspondences, and accordingly slower access times.

To set gram_size:

generator = 7)
# Will take a while the first time, while the database is created.
puts generator.word 


By default, MarkovWords will use the system dictionary located (on Macs) in /usr/share/dict/words. You can change this setting:

# eg to generate random proper names instead of English-sounding words.
generator = '/usr/share/dict/propernames')

This is pretty great, because it means that if you have a dictionary to emulate, you can make words that sound like anything!

Data Storage

MarkovWords stores its database of n-gram concurrences on disk and loads it into memory when necessary. You can control the location of the data file with:

# eg to store the data in /tmp/
generator = /tmp/

You can also clear out the contents of the data file (because MarkovWords will re-use it by default), by passing flush_data: true:

# eg to store the data in /tmp/
generator = /tmp/, flush_data: true)

Custom Metadata

A Generator object gives you access to its .data_store, which is an instance of a FileStore object. This gives you the ability to store custom metadata into the same database that holds the n-gram information.

One example of how you might use this would be to cache words for later use (since initial word generation can be slow, even after the database has been generated the first time):

generator =
my_cache = { generator.word }
generator.data_store.store_data :cache, my_cache

# then later, perhaps on another page load in a web server...
my_cache = generator.data_store.retrieve_data :cache


We've included a bin/benchmark script, which will measure initial load times, and then the time it takes to generate 100 words at various dictionary n-gram sizes.

Here is an example run:

bin/benchmark 1 6 '/usr/share/dict/words'
Minimum n-gram size set to 1
Maximum n-gram size set to 6
Corpus file set to /usr/share/dict/words

Test initial database creation time versus gram size? (y/n) y
user     system      total        real
size: 1   4.080000   0.010000   4.090000 (  4.108898)
size: 2   8.320000   0.090000   8.410000 (  8.554122)
size: 3  12.710000   0.080000  12.790000 ( 12.869257)
size: 4  18.750000   0.160000  18.910000 ( 19.102232)
size: 5  25.440000   0.250000  25.690000 ( 25.953532)
size: 6  31.060000   0.340000  31.400000 ( 31.680680)

Test existing database on disk, initial memory load? (y/n) y
user     system      total        real
size: 1   0.000000   0.000000   0.000000 (  0.000587)
size: 2   0.000000   0.000000   0.000000 (  0.005109)
size: 3   0.080000   0.010000   0.090000 (  0.077303)
size: 4   0.330000   0.070000   0.400000 (  0.395079)
size: 5   1.030000   0.130000   1.160000 (  1.157014)
size: 6   2.920000   0.120000   3.040000 (  3.045219)

Test word generation averages for 100 words per gram size? (y/n) y
user     system      total        real
size: 1   0.010000   0.000000   0.010000 (  0.003971)
size: 2   0.010000   0.000000   0.010000 (  0.009460)
size: 3   0.120000   0.000000   0.120000 (  0.127297)
size: 4   0.350000   0.010000   0.360000 (  0.354564)
size: 5   2.250000   0.020000   2.270000 (  2.302405)
size: 6   4.000000   0.120000   4.120000 (  4.186757)

Change Log

  • 2.0.0
    • Breaking changes:
      • Removed all caching functions from Generator. They were cluttering up the code, without being a necessary function of a Generator.
      • Added an attr_accessor for Generator.data_store, so that users can implement custom metadata for Generator objects, and store it in the same FileStore object that holds the database.
  • 1.0.0 introduced a couple of breaking changes:
    • Words class renamed to Generator.
    • Generator:
      • cache: [boolean] parameter was re-named to perform_caching: [boolean].
      • Removed a lot of attr_accessor variables such as data_store, min_length, max_length etc., in favor of a leaner + cleaner API.
      • The cache file is no longer persisted to disk separately (because FileStore is using SQLite instead of direct-disk storage).
  • 0.2.x was all about Rubocop compliance, so it was a few method refactors but nothing major.
  • 0.1.0 initial commit


After checking out the repo, run bin/setup to install dependencies. Then, run rake test to run the tests. You can also run bin/console for an interactive prompt that will allow you to experiment.

To install this gem onto your local machine, run bundle exec rake install. To release a new version, update the version number in version.rb, and then run bundle exec rake release, which will create a git tag for the version, push git commits and tags, and push the .gem file to


Bug reports and pull requests are welcome on GitHub at This project is intended to be a safe, welcoming space for collaboration, and contributors are expected to adhere to the Contributor Covenant code of conduct.


The gem is available as open source under the terms of the MIT License.

Code of Conduct

Everyone interacting in the MarkovWords project’s codebases, issue trackers, chat rooms and mailing lists is expected to follow the code of conduct.