AI-couplet-writer

Welcome! This is our fun project of using AI to write couplets (对对联). Please check out our website!

Introduction

What is couplet (对联)
- Couplet is a two-liner poem with strict semantic and phoenetic rules
- See the Wikipedia (English) page for more details
What is this project for
- The model will predict the second line (下联) given the first line (上联)
- We are a Chinese couple living in California and we love traditional Chinese poems!
- This project improves from similar ones (see Reference for their great work) on:
  - This is the first of its kind written in TensorFlow 2, to the best of our knowledge
  - This model improves from the reference models mostly on (i) emotional (意境) matching; (ii) repeated characters treatment (see the Model section below)

This is a sequence-to-sequence model with Encoder + Decoder + Attention, schematically shown as below.

Notable features

Download the training data (~70k couplets) from this repo
Set up the model hyperparameters and the file paths in the baseconfig.ini
Install the package dependencies
- TensorFlow 2.2.0
- Python 3.6.9
- Numpy 1.18.5
- Gensim 3.6.0
- Scikit-Learn 0.22.0
Train the model
- Run as python train.py
- The model parameters will be written to the path set up in the config file
- Our model on the website was trained on Google Colab Pro for 12 epochs (~1 day)