A toy-project for random text generation in Haskell. This is fairly basic stuff and in need of serious optimization and improvements.

The name comes from Frédéric Lordon, a French economist known for his... how shall we put it ?... fairly idiosyncratic style. The purpose of this admittedly silly project is to ultimately build a Twitter bot mimicking Mr. Lordon's rather unique prose.

Of course, this little application can actually be used to generate any kind of random texts, provided you have a training set.

Important note: due to copyright issues, I don't want to put the source material on the repository. So you'll need to provide your own. By default, building will expect a "data/" directory, containing a file named "lordon.txt".

Generate random text by using the lordonator command, or if you use Stack:

stack exec lordonator

The process can be rather slow, particularly if your source text is big.

Though there is automatic --help available, here's a quick overview of available options:

The --source option lets you use another source file than the packaged one. Source file should be flat text.

The --sentences option lets you require a given number of sentences. By default, only one randomish sentence will be generated.

Model training can be set up through --up or -u for short and --down or -o for short. The logic is rather simple : we make a map from n-grams to a probability for every sequence of words known to follow the n-gram. The first number is the "n" of "n-gram"; the second the lenght of the sequence of words. Higher number will yield more coherent result, but have higher chances of being mere quotation of the source. Lower number yield potential hilarious results, but with frequent gibberish.

Finally, setting The --dry flag won't generate anything, but will print the training model. (Which tends to be pretty big, so redirecting your output to a file might be a good idea).

Well currently, not really, no. There are tons of issues to fix. But the result might be funny nonetheless. My todo list includes:

• Reading something serious about Markov algorithm rather than doing this by intuition;

• Improving the normalization of source files, which right now, is kind of a disgrace;

• Finding a way to end sentences in a less abrupt fashion.