yxuansu/HCL

[ACL'21] Dialogue Response Selection with Hierarchical Curriculum Learning

Dialogue Response Selection with Hierarchical Curriculum Learning

Authors: Yixuan Su, Deng Cai, Qingyu Zhou, Zibo Lin, Simon Baker, Yunbo Cao, Shuming Shi, Nigel Collier, and Yan Wang

Code for ACL 2021 paper Dialogue Response Selection with Hierarchical Curriculum Learning

Introduction:

In this repository, we provide a simpler and more robust implementation of our ACL 2021 paper and it requires less hyper-parameter tuning. We provide data, pre-trained models for Douban dataset. We will update data for Ubuntu and E-commerce soon.

1. Enviornment Installtion:

pip install -r requirements.txt

2. Download Data here:

unzip data.zip and replace it with the empty ./data folder.

3. SABERT

(1) GPU Requirement:

a. 4 x Tesla V100 GPUs(16GB)
b. Cuda Version: 11.0

(2) Download pre-trained BERT parameter here:

unzip bert-base-chinese.zip and replace it with the empty ./SABERT/bert-base-chinese folder

(3) Training from scratch:

cd ./SABERT
chmod +x ./train.sh
./train.sh

(4) Inference from pre-trained checkpoints:

(a) Download pre-trained parameters here:

unzip ckpt.zip and replace it with the empty ./SABERT/ckpt folder

(b) Perform inference:

cd ./SABERT
chmod +x ./inference.sh
./inference.sh

4. SMN and MSN

(1) GPU Requirement:

a. 1 x Tesla V100 GPUs(16GB)
b. Cuda Version: 11.0

(2) Download embeddings for both models here:

unzip embeddings.zip and replace it with the empty ./SMN_MSN/embeddings folder

(3) Training from scratch:

cd ./SMN_MSN/
chmod +x train_X.sh (X in ['smn', 'msn])
./train_X.sh

(4) Inference from pre-trained checkpoints:

(a) Download pre-trained parameters here:

unzip ckpt.zip and replace it with the empty ./SMN_MSN/ckpt folder

(b) Perform inference:

cd ./SMN_MSN
chmod +x ./inference_X.sh (X in ['smn', 'msn])
./inference_X.sh

5.Citation

If you find our paper and resources useful, please kindly cite our paper:

@inproceedings{su-etal-2021-dialogue,
    title = "Dialogue Response Selection with Hierarchical Curriculum Learning",
    author = "Su, Yixuan  and
      Cai, Deng  and
      Zhou, Qingyu  and
      Lin, Zibo  and
      Baker, Simon  and
      Cao, Yunbo  and
      Shi, Shuming  and
      Collier, Nigel  and
      Wang, Yan",
    booktitle = "Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)",
    month = aug,
    year = "2021",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.acl-long.137",
    doi = "10.18653/v1/2021.acl-long.137",
    pages = "1740--1751"
}