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

Fine-tuning

Performance

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