Learning from Easy to Complex: Adaptive Multi-curricula Learning for Neural Dialogue Generation

This repo contains preliminary code of the AAAI2020 paper named "Learning from Easy to Complex: Adaptive Multi-curricula Learning for Neural Dialogue Generation".

This codebase is built upon the ParlAI project. Check parlai/agents/adaptive_learning for experimental models implementation. RL-based multi-curriculum learning lies in parlai/tasks/adaptive_learning. Running scripts can be found in projects/adaptive_learning.

Framework Overview

Requirements

• Python3
• Pytorch 1.2 or newer

Dependencies of the core modules are listed in requirement.txt.

Dataset

The datasets used in the paper can be download from here. Put it in data/ and unzip it using tar -xzvf AdaptiveLearning.tar.gz

Installing

git clone git@github.com:hengyicai/Adaptive_Multi-curricula_Learning_for_Dialog.git ~/Adaptive_Multi-curricula_Learning_for_Dialog
cd ~/Adaptive_Multi-curricula_Learning_for_Dialog; python setup.py develop
echo "export PARLAI_HOME=~/Adaptive_Multi-curricula_Learning_for_Dialog" >> ~/.bashrc; source ~/.bashrc

Running

cd ~/Adaptive_Multi-curricula_Learning_for_Dialog
bash projects/adaptive_learning/shell/run.sh

The last line of projects/adaptive_learning/shell/run.sh specifies preliminary arguments for the training:

# train_model  MODEL_NAME  TASK_NAME  SUB_TASK  T  VALIDATION_EVERY_N_SECS  VALIDATION_EVERY_N_EPOCHS  NUM_EPOCHS
train_model seq2seq personachat_h3 combine 11000 -1 0.2 30

This run will apply the multi-curriculum learning framework on Seq2seq model using dataset PersonaChat. The duration of curriculum learning is 11000 steps.

Applying the single specificity curriculum dialogue learning on model CVAE using dataset DailyDialog, with curriculum learning duration 8000:

train_model cvae daily_dialog specificity 8000 -1 0.2 30

See projects/adaptive_learning/shell/run.sh for details.

Citation

If you find our code/models or ideas useful in your research, please consider citing the paper:

@InProceedings{Hengyi_2020_AAAI,
  author={Hengyi Cai and Hongshen Chen and Cheng Zhang and Yonghao Song and Xiaofang Zhao and Yangxi Li and Dongsheng Duan and Dawei Yin},
  title={Learning from Easy to Complex: Adaptive Multi-curricula Learning for Neural Dialogue Generation},
  booktitle = {Proceedings of the Thirty-Fourth AAAI Conference on Artificial Intelligence (AAAI)},
  year = {2020}
}