LeD

"Enhancing Multi-Label Text Classification under Label-Dependent Noise: A Label-Specific Denoising Framework" EMNLP 2024.

Requirements:

• python==3.9.19
• numpy==1.26.4
• scipy==1.13.0
• scikit-learn==1.4.2
• torch==2.3.0
• gensim==4.1.2
• nltk==3.8.1
• tqdm==4.66.2
• tokenizers==0.19.1
• transformers==4.41.2

Datasets

LeD uses the same datasets with nEM.

Experiments

Data Path

Please confirm the corresponding configuration file. Make sure the data path parameters (data_dir, dataset and etc.) are right in:

main.py

Train and Evaluate

bash script/run.sh <dataset> <noise_rate> <gpu_id> 

options:
<dataset>: aapd, rcv, movie, riedel.
<noise_rate>: 0.2, 0.4, 0.6