Are you sure you want to delete this task? Once this task is deleted, it cannot be recovered.
You can not select more than 25 topics
Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.
tal
e5ce6d9b1c
|
1 year ago | |
|---|---|---|
| data | 1 year ago | |
| src | 1 year ago | |
| 3500_frequent_chars.txt | 1 year ago | |
| README.md | 1 year ago | |
| synonym.json | 1 year ago | |
README.md
Source code and data for Robust Learning for Text Classification with Multisource Noise Simulation and Hard Example Mining.
How to run:
step1: run scripts in noise_simulation folder to generate simultaed samples
step2: python mining.py with the correct data path
step3: python inference.py
Dependency:
pytorch>=1.5.0
python>=3.5.5
Numpy
该算法提出了一个新颖的、鲁棒的训练框架，可以提升模型在基于 OCR 识别文本的下游任务表现。该框架：1）采用简单而有效的方法，从干净的文本中直接模拟自然的 OCR 噪音，2）从大量的模拟样本中反复挖掘难样本，以获得最佳性能。3）采用稳定性损失，使模型学习到不受噪音表征的影响。在 Metaphor 数据集上 accuracy 达到 87.7%，F1 达到 87.1%。
CSV Text Python
