hanjanghoon/BERT_FP

Fine-grained Post-training for Improving Retrieval-based Dialogue Systems - NAACL 2021

Fine-grained Post-training for Multi-turn Response Selection

Implements the model described in the following paper Fine-grained Post-training for Improving Retrieval-based Dialogue Systems in NAACL-2021.

@inproceedings{han-etal-2021-fine,
title = "Fine-grained Post-training for Improving Retrieval-based Dialogue Systems",
author = "Han, Janghoon  and Hong, Taesuk  and Kim, Byoungjae  and Ko, Youngjoong  and Seo, Jungyun",
booktitle = "Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
month = jun, year = "2021", address = "Online", publisher = "Association for Computational Linguistics", url = "https://www.aclweb.org/anthology/2021.naacl-main.122", pages = "1549--1558",
}

This code is reimplemented as a fork of huggingface/transformers.

Setup and Dependencies

This code is implemented using PyTorch v1.8.0, and provides out of the box support with CUDA 11.2 Anaconda is the recommended to set up this codebase.

# https://pytorch.org
conda install pytorch==1.8.0 torchvision==0.9.0 torchaudio==0.8.0 cudatoolkit=11.1 -c pytorch -c conda-forge
pip install -r requirements.txt

Preparing Data and Checkpoints

Post-trained and fine-tuned Checkpoints

We provide following post-trained and fine-tuned checkpoints.

• fine-grained post-trained checkpoint for 3 benchmark datasets (ubuntu, douban, e-commerce)
• fine-tuned checkpoint for 3 benchmark datasets (ubuntu, douban, e-commerce)

Data pkl for Fine-tuning (Response Selection)

We used the following data for post-training and fine-tuning

• fine-grained post-training dataset and fine-tuning dataset for 3 benchmarks (ubuntu, douban, e-commerce)

Original version for each dataset is availble in Ubuntu Corpus V1, Douban Corpus, and E-Commerce Corpus, respectively.

Fine-grained Post-Training

Making Data for post-training and fine-tuning

Data_processing.py

Post-training Examples

(Ubuntu Corpus V1, Douban Corpus, E-commerce Corpus)

python -u FPT/ubuntu_final.py --num_train_epochs 25
python -u FPT/douban_final.py --num_train_epochs 27
python -u FPT/e_commmerce_final.py --num_train_epochs 34

Fine-tuning Examples

(Ubuntu Corpus V1, Douban Corpus, E-commerce Corpus)

Taining

To train the model, set `--is_training`
python -u Fine-Tuning/Response_selection.py --task ubuntu --is_training
python -u Fine-Tuning/Response_selection.py --task douban --is_training
python -u Fine-Tuning/Response_selection.py --task e_commerce --is_training

Testing

python -u Fine-Tuning/Response_selection.py --task ubuntu
python -u Fine-Tuning/Response_selection.py --task douban 
python -u Fine-Tuning/Response_selection.py --task e_commerce

Training Response Selection Models

Model Arguments

Fine-grained post-training

task_name	data_dir	checkpoint_path
ubuntu	ubuntu_data/ubuntu_post_train.pkl	FPT/PT_checkpoint/ubuntu/bert.pt
douban	douban_data/douban_post_train.pkl	FPT/PT_checkpoint/douban/bert.pt
e-commerce	e_commerce_data/e_commerce_post_train.pkl	FPT/PT_checkpoint/e_commerce/bert.pt

Fine-tuning

task_name	data_dir	checkpoint_path
ubuntu	ubuntu_data/ubuntu_dataset_1M.pkl	Fine-Tuning/FT_checkpoint/ubuntu.0.pt
douban	douban_data/douban_dataset_1M.pkl	Fine-Tuning/FT_checkpoint/douban.0.pt
e-commerce	e_commerce_data/e_commerce_dataset_1M.pkl	Fine-Tuning/FT_checkpoint/e_commerce.0.pt

Performance

We provide model checkpoints of BERT_FP, which obtained new state-of-the-art, for each dataset.

Ubuntu	R@1	R@2	R@5
[BERT_FP]	0.911	0.962	0.994

Douban	MAP	MRR	P@1	R@1	R@2	R@5
[BERT_FP]	0.644	0.680	0.512	0.324	0.542	0.870

E-Commerce	R@1	R@2	R@5
[BERT_FP]	0.870	0.956	0.993