CGI

This repository contains data and code for the thesis "Contrastive Goal Inference for Dialogue Systems".

This study presents a new contrastive reinforcement learning for multi-domain task-oriented dialogue based on the proposed Contrastive Goal Inference (CGI) where contrastive learning is performed and leveraged without the needs of auxiliary perception loss for data augmentation. The CGI is realized in a reinforcement learning algorithm to learn an informative goal from knowledge-based.

The demonstration video: https://youtu.be/Km6ABG0QspM

Requirements

You can directly use docker:

sudo docker run -it --rm --shm-size 8G --privileged --gpus all chin0880ee/cgi bash

and git clone this repository.

The implementation is based on Python 3.10.4 To install the dependencies used in this project. If you don't use docker, please run:

pip install nltk scikit-learn tqdm matplotlib seaborn line-bot-sdk flask torch==1.13.0

Relative environment: Ubuntu==22.04, Python==3.10.6, nvidia driver==470.141.03

Quickstart

Step 1: Training

For different datasets, please first set up the following parameters in the script run_train.sh:

data_dir=[xxx]        # directory of the specific dataset
save_dir=[xxx]        # directory to store trained models

and then run:

bash run_train_[smd or woz].sh

Note that more arguments for training can be found in the main.py.

For self-critical sequence training, please set up num_epochs larger than pre_epochs but no larger than 3 epochs, since self-critical sequence training for a long time might be unstable sometimes.

Step 2: Testing

For different datasets, please first set up the following parameters in the script run_test.sh:

data_dir=[xxx]        # directory of the specific dataset
save_dir=[xxx]        # directory of the trained models
output_dir=[xxx]      # directory to store generation results

and then run:

bash run_test_[smd or woz].sh

Note that more arguments for testing can be found in the main.py.

Step 3: Evaluation

For different datasets, please first set up the following parameters in the script run_eval.sh:

data_name=[xxx]      # ['smd', 'multiwoz']
data_dir=[xxx]       # directory of the specific dataset
eval_dir=[xxx]       # directory of the generation output

and then run:

bash run_eval_[smd or woz].sh

Result

MultiWOZ		SMD
BLEU	Ent. F1	BLEU	Ent. F1
14.5	26.9	18.0	63.3

Acknowledgement

Dual Dynamic Memory Network for End-to-End Multi-turn Task-oriented Dialog Systems. Jian Wang, Junhao Liu, Wei Bi, Xiaojiang Liu, Kejing He, Ruifeng Xu, Min Yang. COLING 2020. [PDF] [Code]

We are highly grateful for the public code of DDMN!