xue2han/poetry-seq2seq

Chinese Poetry Generation

Chinese Poetry Generation

This project aims to implement and improve upon the classical Chinese poetry generation system proposed in "Chinese Poetry Generation with Planning based Neural Network".

Generated Sample

Dependencies

Python 2.7
TensorFlow 1.8.0
Jieba 0.38
Gensim 2.0.0
pypinyin 0.23

Features

Network:

• Bidirectional encoder
• Attention decoder

Training and Predicting:

• Alignment boosted word2vec
• Data loading mode: only keywords (no preceding sentences)
• Data loading mode: reversed
• Data loading mode: aligned
• Training mode: ground truth
• Training mode: scheduled sampling
• Predicting mode: greedy
• Predicting mode: sampling
• Predicting mode: beam search

Refinement:

• Output refiner
• Reinforcement learning tuner
• Iterative polishing

Evaluation:

• Evaluation: rhyming
• Evaluation: tonal structure
• Evaluation: alignment score
• Evaluation: BLEU score

Project Structure

Data
data: directory for raw data, processed data, pre-processed starterkit data, and generated poetry samples
model: directory for saved neural network models
log: directory for training logs
notebooks: directory for exploratory/experimental IPython notebooks
training_scripts: directory for sample scripts used for training several basic models

Code
model.py: graph definition
train.py: training logic
predict.py: prediction logic
plan.py: keyword planning logic
main.py: user interaction program

Data Processing

To prepare training data:

python data_utils.py

Detail
This scrip does the following in order:

1. Parse corpus
2. Build vocab
3. Filter quatrains
4. Count words
5. Rank words
6. Generate training data

Note
The TextRank algorithm may take many hours to run.
Instead, you can choose to interrupt the iterations and stop it early,
when the progress shown in the terminal has remained stationary for a long time.

Then, to generate the word embedding:

python word2vec.py

Alternative
As an alternative, we have also provided pre-processed data in the data/starterkit directory
You may simply perform cp data/starterkit/* data/processed to skip the data processing step

Training

To train the default model:

python train.py

To view the full list of configurable training parameters:

python train.py -h

Note
Thus you should almost always train a new model after modifying any of the parameters.
Models are by default saved to model/. To train a new model, you may either remove the existing model from model/
or specify a new model path during training with python train.py --model_dir :new_model:dir:

Generating

To start the user interation program:

python main.py

Similarly, to view the full list of configurable predicting parameters:

python main.py -h

Note
The program currently does not check that predication parameters matches corresponding training parameters.
User has to ensure, in particular, the data loading modes correspond with the ones used during traing.
(e.g. If training data is reversed and aligned, then prediction input should also be reversed and aligned.
Otherwise, results may range from subtle differences in output to total crash.

Evaluating

To generate sample poems for evaluation:

python generate_samples.py

Detail
This script by default randomly samples 4000 poems from the training data and saves them as human poems. Then it uses entire poems as inputs to the planner, to create keywords for the predictor. The predicted poems are saved as machine poems.

To evaluate the generated poems:

python evaluate.py

Further Reading

Auxiliary

1. "Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks"
2. "Sequence-to-Sequence Learning as Beam-Search Optimization"
3. "Tuning Recurrent Neural Networks with Reinforcement Learning"
4. "Deep Reinforcement Learning for Dialogue Generation"

Poetry Generation

1. May 10, 2017: "Flexible and Creative Chinese Poetry Generation Using Neural Memory"
2. Dec 7, 2016: "Chinese Poetry Generation with Planning based Neural Network"
3. June 19, 2016: "Can Machine Generate Traditional Chinese Poetry? A Feigenbaum Test"

Acknowledgement

1. The data processing source code is based on DevinZ1993's implementation.
2. The neural network implementation is inspired by JayParks's work.