1429904852/ldf

[Findings of EMNLP 2022] Label-Driven Denoising Framework for Multi-Label Few-Shot Aspect Category Detection

LDF

Code and data for "Label-Driven Denoising Framework for Multi-Label Few-Shot Aspect Category Detection" (Findings of EMNLP 2022)

Overview

• In this paper, we propose a Label-Driven Denoising Framework (LDF) to alleviate the noise problems for the FS-ACD task.
• Label-Driven Denoising Framework contains a label-guided attention strategy to filter noisy words and generate a representative prototype for each aspect, and a label-weighted contrastive loss to avoid generating similar prototypes for semantically-close aspect categories.

Setup

Requirements

+ python 3.7
+ tensorflow 2.4.0
+ keras 2.4.3
+ sklearn 0.0
+ numpy 1.19.5

Download word embedding

please download the glove.6B.50d embedding in JSON format: [Link], or in txt format: [StanfordNLP] and put it under word_embedding folder

Model configuration

• you can choose one or multiple methods at one time in the model_list

e.g., model_list = [None, 'AWATT_LAS', 'LDF_AWATT']

# code:             corresponding model:
#  None             the original AWATT model
# 'AWATT_LAS'       AWATT+LAS
# 'AWATT_LCL'       AWATT+LCL
# 'AWATT_SCL'       AWATT+SCL
# 'LDF_AWATT'       LDF-AWATT
# 'HATT'            the original HATT model
# 'HATT_LAS'        HATT+LAS
# 'HATT_LCL'        HATT+LCL
# 'HATT_SCL'        HATT+SCL
# 'LDF-HATT'        LDF-HATT

• you can choose one or multiple datasets at one time in the dataset_list

e.g., dataset_list = ['FewAsp', 'FewAsp(single)', 'FewAsp(multi)']

• you can choose one or multiple configs at one time in the config_list

e.g., config_list = [[2, 5, 5], [1, 5, 10], [1, 10, 5], [1, 10, 10]]

# [2, 5, 5] stands for: two(2) '5'-way-'5'-shot meta-tasks for two batch-size
# [1, 5, 10] stands for: one(1) '5'-way-'10'-shot meta-task for one batch-size
# [1, 10, 5] stands for: one(1) '10'-way-'5'-shot meta-task for one batch-size
# [1, 10, 10] stands for: one(1) '10'-way-'10'-shot meta-task for one batch-size

Usage

• You can use the following command to train and test LDF on the FS-ACD task:

python train_and_test.py

• The final results can be saved in the excel file you specified:

e.g., pd.DataFrame(result_list).to_excel('result.xlsx')

Implementation details

• The implementation of Label-weighted Contrastive Loss follows the simplification below:

• For the numeric results in the experiments, we take 5 runs covering seeds [5, 10, 15, 20, 25]. Different GPUs and versions of Keras/TensorFlow might give different results. Feel free to use our code, re-implement, and re-run the experiments!

• For the implementation of model [AWATT] whose code is not available when we are working on LDF, in order to achieve the reported results, our implementation slightly differs from what is described in [paper].

Build your own model

you can augment your own model with LDF by:

• Introduce label text into the Attention module to help focus on salient information that benefits classification;

• Add our Label-weighted Contrastive Loss.

Citation

If the code is used in your research, please cite the paper:

@inproceedings{zhao-etal-2022-label,
    title = "Label-Driven Denoising Framework for Multi-Label Few-Shot Aspect Category Detection",
    author = "Zhao, Fei  and
      Shen, Yuchen  and
      Wu, Zhen  and
      Dai, Xinyu",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2022",
    month = dec,
    year = "2022",
    address = "Abu Dhabi, United Arab Emirates",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.findings-emnlp.177",
    pages = "2390--2402"
}

If the data is used in your research, please cite the paper:

@inproceedings{hu-etal-2021-multi-label,
    title = "Multi-Label Few-Shot Learning for Aspect Category Detection",
    author = "Hu, Mengting and Zhao, Shiwan and Guo, Honglei and Xue, Chao and Gao, Hang and Gao, Tiegang and Cheng, Renhong and Su, Zhong",
    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.495",
    doi = "10.18653/v1/2021.acl-long.495",
    pages = "6330--6340",
}

ldf's Issues

关于论文《Multi-Label Few-Shot Learning for Aspect Category Detection》

不知道作者有没有此篇论文代码，因为我尝试用pytorch复现LDF，结果差距还挺大的，我排除了许多因素还是没有头绪。 QAQ

关于训练过程的标准

有几个疑惑没想明白：
1、采用多少个epochs，每个epoch又分task训练800次，验证600次吗？
2、5-way、5-shot中：每个epoch随机选取5 * 5个suppor和5个query（也就是总共30条数据）吗？还是每个task随机选取5 * 5个suppor和5个query吗（也就是总共800 * 30+600 * 30条
3、5-way、5-shot中两个batch是分别取两次随机数据吗（也就是总共60条数据）？

数据）？

关于softmax的问题

最后我们会得到每个样本对每个原型的距离：
那么我们softmax是对一个样本对每个原型的距离，
还是说每个样本对其中一个原型的softmax。