Markov
Text generation library based on nth-order Markov chains
Features
- Token sanitation (optional): ignores letter case and punctuation when switching states, but still keeps the output as-is
- Operation history (optional): recalls the operations it was instructed to perform, incl. past training data
- Probability shifting (optional): gives less popular generation paths more chance to get used, which makes the output more original but may produce nonsense
- Tagging (optional): you can tag your source data to be queried later by aggregating those tags in any way you want, kind of like a database
- Context awareness (optional) grants your model the ability to answer questions given to it provided training data is good enough
- Managed disk storage
- Transparent fragmentation reduces RAM usage and loading times with huge models
Usage
In mix.exs:
defp deps do
[{:markov, "~> 3.0"}]
end
Before using for the first time:
$ mix amnesia.create -d Markov.Database --disk
Unlike Markov 1.x, this version has very strong opinions on how you should create and persist your models (that differs from 2.x).
Example workflow (click here for full docs):
# The name can be an arbitrary term (not just a string).
# It will be stored in a Mnesia DB and created from scratch using the specified
# parameters if not found.
# You should configure mnesia if you want to change its working dir, e.g.:
# `config :mnesia, dir: "/var/data"`
{:ok, model} = Markov.load("model_name", sanitize_tokens: true, store_log: [:train])
# train using four strings
:ok = Markov.train(model, "hello, world!")
:ok = Markov.train(model, "example string number two")
:ok = Markov.train(model, "hello, Elixir!")
:ok = Markov.train(model, "fourth string")
# generate text
{:ok, text} = Markov.generate_text(model)
IO.puts(text)
# commit all changes and unload
Markov.unload(model)
# these will return errors because the model is unloaded
# Markov.generate_text(model)
# Markov.train(model, "hello, world!")
# load the model again
{:ok, model} = Markov.load("/base/directory", "model_name")
# enable probability shifting and generate text
:ok = Markov.configure(model, shift_probabilities: true)
{:ok, text} = Markov.generate_text(model)
IO.puts(text)
# print uninteresting stats
model |> Markov.dump_partition(0) |> IO.inspect
model |> Markov.read_log |> IO.inspect
# this will also write our new just-set option
Markov.unload(model)