FATRER

[ECAI 2023] Official Pytorch implementation of "FATRER: Full-Attention Topic Regularizer for Accurate and Robust Conversational Emotion Recognition" [paper]

Framework

Full-attention topic regularizer(FATRER) introduces an emotion-related global view when modeling the local context in a conversation. A joint topic modeling strategy is introduced to implement regularization from both representation and loss perspectives. To avoid overregularization, FATRER drops the constraints on prior distributions that exist in traditional topic modeling and perform probabilistic approximations based entirely on attention alignment. Experiments show that FATRER obtain more favorable results than state-of-the-art models, and gain convincing robustness.

News

• [2023-10-02]: FARTER will be presented orally(Video) in ECAI2023(Programme) Technical Session 1 at 09:30 AM in Room S4A and Poster Session 1 at 11:15 AM in Hall S3B(Poster).
• [2023-07-15]: FARTER has been accepted by ECAI 2023 (Paper 223).

Prerequisites

• Python 3.9.12
• Pytorch 1.10.1+cu113

  pip instll -r requirements.txt

Usage

Benchmark Datasets

• IEMOCAP/MELD/EmoryNLP/EmoryNLP

Generalization results on four datasets

Execution

IEMOCAP

1. FARTER-Multi:

    # train
    python main.py conf/FATRER_multi.yaml

    #train and conduct attack(U+C) based on PWWS(per 50 epoch):
    python main.py conf/FATRER_multi_pwws_attack.yaml

    #train and conduct attack(U+C) based on TextFooler(per 50 epoch):
    python main.py conf/FATRER_multi_textfooler_attack.yaml

    #train and conduct attack(U+C) based on TextBugger(per 50 epoch):
    python main.py conf/FATRER_multi_textbugger_attack.yaml

2. FARTER-Multi(without topic-oriented regularization):

    # train
    python main.py conf/FATRER_multi_wo_topic.yaml

3. FARTER-Single:

    # train
    python main.py conf/FATRER_single.yaml

4. FARTER-Single(without topic-oriented regularization):

    #train
    python main.py conf/FATRER_single_wo_topic.yaml

5. DialTRM(Baseline):

    #train
    python main.py conf/Baseline.yaml

6. VAE(topic-oriented)

  #train VAE(Laplace)
  python main.py conf/VAE_Laplace.yaml

  #train VAE(Dirichlet)
  python main.py conf/VAE_Dirichlet.yaml

  #train VAE(Gamma)
  python main.py conf/VAE_Gamma.yaml

  #train VAE(LogNormal)
  python main.py conf/VAE_LogNormal.yaml

MELD

1. FARTER-Multi:

    python main.py conf/FATRER_multi_MELD.yaml

2. FARTER-Single:

    python main.py conf/FATRER_single_MELD.yaml

EmoryNLP

1. FARTER-Multi:

    python main.py conf/FATRER_multi_EmoryNLP.yaml

2. FARTER-Single:

    python main.py conf/FATRER_single_EmoryNLP.yaml

DailyDialog

1. FARTER-Multi:

    python main.py conf/FATRER_multi_DailyDialog.yaml

2. FARTER-Single:

    python main.py conf/FATRER_single_DailyDialog.yaml

Cite us

Please cite the following paper if you find this code useful in your work.

@article{mao2023fatrer,
  title={FATRER: Full-Attention Topic Regularizer for Accurate and Robust Conversational Emotion Recognition},
  author={Mao, Yuzhao and Lu, Di and Wang, Xiaojie and Zhang, Yang},
  journal={arXiv preprint arXiv:2307.12221},
  year={2023}
}

License

MIT license