Official codebase for the paper "Structural Contrastive Representation Learning for Zero-shot Multi-label Text Classification" in Findings of EMNLP 2022
Code tested with Python 3.7
on Ubuntu 20.04
.
To run, please first install the packages in requirements.txt
and follow the instructions in dataset/README.md
to download datasets.
# LF-Amazon-131K
python train.py --dataset dataset/LF-Amazon-131K --min-len 40 --max-len 80 --batch-size 384 --experiment a131k
# LF-Amazon-1M
python train.py --dataset dataset/LF-Amazon-1M --min-len 40 --max-len 80 --batch-size 384 --lr 5e-6 --min-lr 5e-7 --experiment a1m
# LF-WikiSeeAlso-320K
python train.py --dataset dataset/LF-WikiSeeAlso-320K --min-len 80 --max-len 160 --batch-size 256 --lr 5e-8 --min-lr 5e-9 --epochs 5 --experiment wsa320k
# LF-Wikipedia-500K
python train.py --dataset dataset/LF-Wikipedia-500K --min-len 80 --max-len 160 --batch-size 256 --lr 5e-8 --min-lr 5e-9 --epochs 5 --experiment w500k