Requirements

• python2.7
• tensorflow 0.12
  • CPU: pip install tensorflow==0.12.0
  • GPU: pip install tensorflow-gpu==0.12.0
• nltk, cmudict and stopwords
  • import nltk; nltk.download("cmudict"); nltk.download("stopwords")
• gensim
  • pip install gensim
• sklearn
  • pip install sklearn
• numpy
  • pip install numpy

Data / Models

• datasets/gutenberg/data.tgz: sonnet data, with train/valid/test splits
• datasets/gutenberg/pretrain.tgz: general poetry data ("Background" corpus in section 3 of the paper) used for pretraining word2vec (34M word tokens)
• pretrain_word2vec/dim100/*: pre-trained word2vec model
• trained_model/model.tgz: trained sonnet model

Pre-training Word Embeddings

• The pre-trained word2vec model has already been supplied: pretrain_word2vec/dim100/*
• It was trained on 34M Gutenberg poetry data (in datasets/gutenberg/pretrain.tgz)
• If you want to train your own word embeddings, you can use the python script (uses gensim's word2vec)
  • python pretrain_word2vec.py

Training the Sonnet Model

1. Extract the data; it should produce the train/valid/test splits
   • cd datasets/gutenberg; tar -xvzf data.tgz
2. Unzip the pre-trained word2vec model
   • gunzip pretrain_word2vec/dim100/*
3. Set up model hyper-parameters and other settings, which are all defined in config.py
   • the default configuration is the optimal configuration used in the paper (documented here)
4. Run python sonnet_train.py
   • takes about 2-3 hours on a single K80 GPU to train 30 epochs

Generating Sonnet Quatrain

1. Extract the trained model
   • cd trained_model; tar -xvzf model.tgz
2. Run python sonnet_gen.py -m trained_model
   • the default configuration is the generation configuration used in the paper
   • takes about a minute to generate one quatrain on CPU (GPU not necessary)

usage: sonnet_gen.py [-h] -m MODEL_DIR [-n NUM_SAMPLES] [-r RM_THRESHOLD]
                     [-s SENT_SAMPLE] [-a TEMP_MIN] [-b TEMP_MAX] [-d SEED]
                     [-v] [-p SAVE_PICKLE]

Loads a trained model to do generation

optional arguments:
  -h, --help            show this help message and exit
  -m MODEL_DIR, --model-dir MODEL_DIR
                        directory of the saved model
  -n NUM_SAMPLES, --num-samples NUM_SAMPLES
                        number of quatrains to generate (default=1)
  -r RM_THRESHOLD, --rm-threshold RM_THRESHOLD
                        rhyme cosine similarity threshold (0=off; default=0.9)
  -s SENT_SAMPLE, --sent-sample SENT_SAMPLE
                        number of sentences to sample from using pentameter
                        loss as sample probability (1=turn off sampling;
                        default=10)
  -a TEMP_MIN, --temp-min TEMP_MIN
                        minimum temperature for word sampling (default=0.6)
  -b TEMP_MAX, --temp-max TEMP_MAX
                        maximum temperature for word sampling (default=0.8)
  -d SEED, --seed SEED  seed for generation (default=1)
  -v, --verbose         increase output verbosity
  -p SAVE_PICKLE, --save-pickle SAVE_PICKLE
                        save samples in a pickle (list of quatrains)

Generated Quatrains:

python sonnet_gen.py -m trained_model/ -d 1

Temperature = 0.6 - 0.8
  01  [0.43]  with joyous gambols gay and still array
  02  [0.44]  no longer when he twas, while in his day
  03  [0.00]  at first to pass in all delightful ways
  04  [0.40]  around him, charming and of all his days

python sonnet_gen.py -m trained_model/ -d 2

Temperature = 0.6 - 0.8
  01  [0.44]  shall i behold him in his cloudy state
  02  [0.00]  for just but tempteth me to stop and pray
  03  [0.00]  a cry: if it will drag me, find no way
  04  [0.40]  from pardon to him, who will stand and wait

Crowdflower and Expert Evaluation

• Annotations can be found in the folder: evaluation_annotation/

Media Coverage

• New Scientist (or view the full article here)
• Times UK (or view the full article without subscription here)
• Daily Mail
• Digital Trends
• NVIDIA
• ABC Central News Now, Texas Affiliate
• InfoSurHoy
• BBC Radio 4 (from 1:45:20 to 1:49:20) (or download extracted segment here)
• Replubbica (Italian)
• Datanami
• TechXplore
• RankRed
• EndGadget
• The New York Post
• Dazed
• The Poetry Society
• IEEE Spectrum
• Passw0rd, Creativity and AI (starts at around the 43 minute mark)

Publication

Jey Han Lau, Trevor Cohn, Timothy Baldwin, Julian Brooke and Adam Hammond (2018). Deep-speare: A joint neural model of poetic language, meter and rhyme (Supplementary Material). In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (ACL 2018), Melbourne, Australia, pp. 1948--1958.

Talk

Creativity, Machine and Poetry for a public forum on language [video]