dial2vec

Introduction

This is the repository of our paper 'dial2vec: Self-Guided Contrastive Learning of Unsupervised Dialogue Embeddings'.

Quick Start

We prepare a set of shell scripts for convenient usage.

Model Options: bert,roberta,t5,todbert,blender,plato

Dataset Options: bitod,doc2dial,metalwoz,mwoz,selfdialog,sgd

Installation

git clone https://github.com/AlibabaResearch/DAMO-ConvAI/tree/main/dial2vec
cd dial2vec

# conda create -n dial2vec python=3.8
pip3 -r install requirements.txt

## download all datasets and move them to ./dial2vec/datasets/
## url -> https://drive.google.com/file/d/1KpxQGXg9gvH-2u21bAMykL5N-tpYU2Dr/view?usp=sharing

## download all trained checkpoints and move them to ./dial2vec/output/
## url -> https://drive.google.com/file/d/1JVod0OLyiVeIRVxvA-uk1TKa_zMn-sZK/view?usp=sharing

## download useful PLM and move them to ./dial2vec/model/
## url -> https://drive.google.com/file/d/1Xq_nj-le_Mm6iUUHjltPtJZYd6gmSvNb/view?usp=sharing

Train

Train dial2vec on ${dataset}, we can use the following script:

sh scripts/train/run_plato.sh 'doc2dial' 'train' 5 100  # train dial2vec on doc2dial dataset

Inference

Evaluate the performance of ${model} on ${dataset}, we can use the following script:

sh scripts/inference/run_roberta.sh 'mwoz'  # evaluate roberta on MultiWOZ dataset.

Quick inference

If you want to inference several models on one dataset in one line, then you can receive results for all models.

sh scripts/inference/run_dataset.sh 'bitod'  # evaluate all models including dial2vec and baselines on BiTOD dataset.

If you want to inference one model on several datasets in one line, then you can receive results for all datasets.

sh scripts/inference/run_model.sh 'todbert'   # evaluate TOD-BERT on six datasets.

Citation