Lyrics

Introduction

• Lyrics generation using machine learning techniques.
• FIT 4-508, Tsinghua University.

Setup

• You should add this 'Lyrics' folder to python path.
• Add the following line to your ~/.bashrc or ~/.profile

export PYTHONPATH=$PYTHONPATH:<path_to_Lyrics>

• Then setup the dependencies.
• Anaconda is highly recommended.

conda env create -f environment.yml
source activate lyrics

• Meanwhile, you can pull the data and model which will be used.
• Tip: Dropbox is blocked by the GFW. Make sure your terminal can do the work.
• If not, peek into saved/pull.sh and manually download them in your browser and place them under saved/.

cd saved
bash pull.sh

References

• Survey & Guide

  • Stanford 2017, Neural Text Generation: A Practical Guide
  • Survey of the State of the Art in Natural Language Generation: Core tasks, applications and evaluation
  • Deep Learning for NLG

• Style Transfer

  • AAAI 2018, Style Transfer in Text: Exploration and Evaluation
  • ACL 2017, Hafez: an Interactive Poetry Generation System
  • NIPS 2017 Workshop, Improved Neural Text Attribute Transfer with Non-parallel Data
  • EMNLP 2016, Stylistic Transfer in Natural Language Generation Systems Using Recurrent Neural Networks

• Thematic & Coherent

  • COLING 2016, Chinese Poetry Generation with Planning based Neural Network
  • ICML 2017, Toward Controlled Generation of Text
  • EMNLP 2016, Globally Coherent Text Generation with Neural Checklist Models

• Refining & Editing

  • NIPS 2017, Deliberation networks: Sequence generation beyond one-pass decoding
  • Generating Sentences by Editing Prototypes

• VAE

  • Tutorial on Variational Autoencoders
  • Google, Generating Sentences from a Continuous Space
  • ACL 2017, A Conditional Variational Framework for Dialog Generation
  • Improved Variational Autoencoders for Text Modeling using Dilated Convolutions
  • A Hybrid Convolutional Variational Autoencoder for Text Generation
  • A Hierarchical Latent Variable Encoder-Decoder Model for Generating Dialogues
  • Generating Thematic Chinese Poetry with Conditional Variational Autoencoder

• Reinforcement Learning

  • BBQ-Networks: Efficient Exploration in Deep Reinforcement Learning for Task-Oriented Dialogue Systems

Subdirectories

experiments

• Jupyter notebooks containing experimental records and study notes are kept here.

docs

• Documentation can be found at Lyrics Documentation

utils

• Utility classes handling data and lyrics.

attribute

• Topic model, keyword extraction, clustering.

saved

• Saved data including training set and trained models are kept here.

app

• App served on Deep Lyrics.
• Frontend in React and backend in Flask.

archived

• Code not in use but may come into use in the future.

seq2seq

• seq2seq implemented in PyTorch.

rhythm

• Use bigram and word vector distance to predict word, constrained by rhythm.

preprocess

• Data cleaning, preprocessing

crawler

• Lyrics crawler.