NLG with hmmlearn

Experimental natural language generation with hidden markov models, using the excellent hmmlearn Python library.

Installing dependencies

Running the programs requires Python (tested on 3.5.2). Install required modules with:

$ pip install -r requirements.txt

Training a model

Example for training a model with 8 hidden states based on given text.

The input text is assumed to have one sentence per line (or segment of comparable size). We recommend cleaning up the text as much as possible, leaving only the essential punctuation (if any).

$ mkdir demo
$ ./train.py -n 8 -o demo/hmm < ../input/input.txt

Generating new text

Generate 20 lines of text, 12 words per line, by simulating the hidden markov model obtained in the previous step:

$ ./gen.py -l 20 -w 12 demo/hmm.builtin.8

For comparison, try generating new text based solely on the frequency distribution of words in the input text. .le and .freqdist files are generated as part of the training step. (Despite the file names, the number of hidden states chosen in the training step does not affect their contents.)

$ ./freq.py -l 20 -w 12 demo/hmm.builtin.8.freqdist demo/hmm.builtin.8.le

Lastly, compare the output with text generated by choosing words from the input text at random:

$ ./rnd.py -l 20 -w 12 demo/hmm.builtin.8.le

License

This project, like hmmlearn, uses the BSD license.