Automatic Poetry Generation from Prosaic Text

v1.0

developed at IRIT, Toulouse

www.timvandecruys.be

(modified by Mark Wolff for install on MacOS)

Introduction

Charles/Sylvia is a system for automatic poetry generation, developed within the MELODI group at IRIT, the research institute for computer science in Toulouse. The system has been trained on billions of words extracted from generic web texts; its rhyming knowledge has been extracted from Wiktionary, and it automatically learned an elementary notion of sense by looking at the context of words.

Sylvia writes in English, while Charles is French.

Examples

2020-07-05 23:42:52 nmfdim 1 (tendresse, joie, bonheur)

je sentais les larmes sur son visage
je le ressens au plus profond de mon cœur
merci de ta tendresse , pour ce partage
je t' aime d' amour , c' est un vrai bonheur

la douceur de tes mots me rend malade
tu es mon coeur , j' aime le silence
tu es ma joie dans mes nuits froides
tu me rappelle des souvenirs d' enfance

                                     - Charles

2020-07-05 23:44:53 nmfdim 13 (sorrow, longing, admiration)

it seemed as though he 'd never had a heart attack
after a moment , a sudden silence filled the room
oh , dear , the man said , his voice almost black
i smiled , admiring the sight of my hands in the bathroom

for a moment , i felt a sense of great pride
taking a deep breath , i roused myself to my feet
i closed my eyes , turning my gaze to the far side
i was restless , eager to see something to eat

                                     - Sylvia

Installation and execution

Clone the git repository:

git clone https://github.com/mbwolff/poetry.git

Create a virtual (python3) environment with all the necessary dependencies. The environment can be installed with the command:

pip install -r requirements.txt

Put the required model files (not included) in directory data
Once installed and model files in place, activate the environment, and run python. A poem can then be written using the following commands (for French):

import charles
p = charles.Poem()
p.write()
p.write(nmfDim=1)

For English, replace charles with sylvia.

Model files

Model files (neural network parameters, rhyme dictionary, NMF model, n-gram model) are not included due to their large file size (2.6GB for French, 3.4GB for English). In order to obtain a copy, send a mail to tim.vandecruys@irit.fr

Dependencies

Pytorch is the most important one; all dependencies are stipulated in the file environment.yml, which can be used to create a suitable Anaconda environment. Note that the poetry generation system heavily relies on the Pytorch version of OpenNMT (https://github.com/OpenNMT/OpenNMT-py), which equally needs to be installed.

Reference

Tim Van de Cruys. 2020. Automatic Poetry Generation from Prosaic Text. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), pp. 2471-2480.

@inproceedings{vandecruys2020automatic,
    title = "Automatic Poetry Generation from Prosaic Text",
    author = "Van de Cruys, Tim",
    booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
    year = "2020",
    publisher = "Association for Computational Linguistics",
    pages = "2471--2480",
}