Description

• This repository reimplemented "ConvBERT: Improving BERT with Span-based Dynamic Convolution (Jiang et al., NeurIPS 2020)" based on awesome works (See Reference).

Requirements

1. python=3.7
2. pytorch==1.12.1
3. sentencepiece==0.1.97
4. setproctitle

Dataset

Pretrain: Kowiki (Korean)

• Download

$ cd web-crawler
$ pip install tqdm
$ pip install pandas
$ pip install bs4
$ pip install wget
$ pip install pymongo
$ python kowiki.py

• Result

겨울이 되어서 날씨가 무척 추워요.
['▁겨울', '이', '▁되어', '서', '▁날', '씨', '가', '▁무', '척', '▁추', '워', '요', '.']
[3217, 3625, 677, 3639, 715, 4073, 3636, 106, 4227, 197, 3993, 3796, 3627]

이번 성탄절은 화이트 크리스마스가 될까요?
['▁이번', '▁성', '탄', '절', '은', '▁화', '이트', '▁크리스', '마', '스가', '▁될', '까', '요', '?']
[2894, 89, 4006, 3961, 3641, 269, 649, 1895, 3698, 711, 1453, 3834, 3796, 4308]

겨울에 감기 조심하시고 행복한 연말 되세요.
['▁겨울', '에', '▁감', '기', '▁조', '심', '하', '시', '고', '▁행', '복', '한', '▁연', '말', '▁되', '세', '요', '.']
[3217, 3628, 193, 3643, 53, 3872, 3633, 3650, 3638, 246, 3911, 3640, 63, 3869, 472, 3719, 3796, 3627]

Finetune: Naver sentiment movie corpus v1.0 (Korean)

• Download

$ cd web-crawler
$ wget https://raw.githubusercontent.com/e9t/nsmc/master/ratings_train.txt
$ wget https://raw.githubusercontent.com/e9t/nsmc/master/ratings_test.txt
$ python naver_movie.py

Run

• define your task (pretrain or finetune)

e.g.,

python convbert.py --task pretrain

Reference

• ConvBERT: Improving BERT with Span-based Dynamic Convolution (Jiang et al., NeurIPS 2020)
  • https://arxiv.org/pdf/2008.02496.pdf
  • https://github.com/huggingface/transformers/blob/main/src/transformers/models/convbert/modeling_convbert.py
  • GELU: https://github.com/huggingface/transformers/blob/main/src/transformers/activations.py
  • config: https://huggingface.co/YituTech/conv-bert-base/resolve/main/config.json
  • config: https://huggingface.co/docs/transformers/main_classes/configuration#configuration
• PAY LESS ATTENTION WITH LIGHTWEIGHT AND DYNAMIC CONVOLUTIONS (Wu et al., ICLR 2019)
  • https://arxiv.org/pdf/1901.10430.pdf
• Korean dataset preprocessing
  • https://paul-hyun.github.io/vocab-with-sentencepiece/
• BERT
  • https://paul-hyun.github.io/bert-01/
  • Pytorch Lightning: https://pytorch-lightning.readthedocs.io/en/stable/common/trainer.html
• Depthwise Convolution
  • https://lena-voita.github.io/nlp_course/models/convolutional.html
  • https://gaussian37.github.io/dl-concept-dwsconv/
  • https://youtu.be/T7o3xvJLuHk
• LightWeight Convolution
  • https://bo-son.github.io/2019/04/05/pay_less_attention/
• Gate Linear Unit (GLU)
  • https://medium.com/deeplearningmadeeasy/glu-gated-linear-unit-21e71cd52081