Generating Poetry with Hidden Markov Models

Introduction

This project implements the Baum-Welch algorithm to train a Hidden Markov Model (HMM) using Shakespearean poetry as a training set. In addition, the trained model is used to generate a sample sonnet. This is accomplished by selecting a seed value from the starting distribution of the model and emitting possible observations while propagating through hidden states.

This project arose out of a mini-project in machine learning for CS 155 at Caltech. The project has since been completely restructured, with all old source code removed. The main motivation was to create a cleaner implementation that utilizes a HMM without any domain specific knowledge of poetry. The result is an implementation that is more representative of the strength of a HMM for generating poems. The contributors to the original project at Caltech were Dryden Bouamalay, Ritwik Anand, and Audrey Huang.

The project uses a context-free grammar (via PLY) for parsing the poems to create training data. This training data is then fed to the Baum-Welch algorithm to train matrices. Finally, these matrices are used to generate a sample poem.

Installation

Install the module with:

$ python setup.py install

For PyEnchant to function correctly it may be necessary to install aspell-en using your distribution's package manager.

Run

After installing the module, the hmm script should be available:

$ hmm -s <states> -t <tolerance>

where <states> is the number of hidden states you wish to use for the model, and <tolerance> is the delta at which training will terminate. For example, the following should completely quickly:

$ hmm -s 10 -t 0.01 

I appreciate any feedback. Thanks!