This repository contains the PyTorch source Code for our paper: DARER: Dual-task Temporal Relational Recurrent Reasoning Network for Joint Dialog Sentiment Classification and Act Recognition.

Bowen Xing and Ivor W. Tsang.

ACL 2022 (Findings).

DARER's Architecture:

Our code relies on Python 3.6 and following libraries:

• transformers==1.1.0
• torch-geometric==1.7.0
• torch==1.5.0
• tqdm==4.60.0
• transformers==3.3.1
• numpy==1.19.2
• scikit-learn==0.24.2

LSTM-based Encoder:

DARER/

    # Mastodon //glove
    python -u main.py -lr 1e-3 -l2 1e-8 -dd dataset/mastodon -hd 128 -mc 2 -dr 0.2 -sn 3
  
    # DailyDialog // glove
    python -u main.py -ne 50 -hd 300 -lr 1e-3 -l2 1e-8 -dd dataset/dailydialogue -rnb 10 -sn 2 -mc 5 -dr 0.5
    # DailyDialog // train random word vector 
    python -u main.py -ne 50 -hd 256 -lr 1e-3 -l2 1e-8 -dd dataset/dailydialogue -sn 1 -mc 1e-05 -dr 0.3 -rw

PTLM(pre-trained language model)-based Encoder:

DARER/pre-trained language model/

    # Mastodon // BERT
    python -u main.py -pm bert -bs 16 -sn 4 -dr 0.3 -hd 768 -l2 0.01 -blr 1e-05 -mc 1
    # Mastodon // RoBERTa
    python -u main.py -pm roberta -bs 16 -sn 4 -dr 0.14 -hd 768 -l2 0.0 -blr 1e-05 -mc 1
    # Mastodon // XLNet
    python -u main.py -pm xlnet -bs 12 -sn 4 -dr 0.2 -hd 256 -l2 0.0 -blr 1e-05 -mc 1

We recommend you search the optimal hyper-parameters on your server to obtain the best performances in your own experiment environment.

If the code is used in your research, please star this repo ^_^ and cite our paper as follows:

@inproceedings{xing-tsang-2022-darer,
    title = "{DARER}: Dual-task Temporal Relational Recurrent Reasoning Network for Joint Dialog Sentiment Classification and Act Recognition",
    author = "Xing, Bowen  and
      Tsang, Ivor",
    booktitle = "Findings of the Association for Computational Linguistics: ACL 2022",
    month = may,
    year = "2022",
    address = "Dublin, Ireland",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.findings-acl.286",
    doi = "10.18653/v1/2022.findings-acl.286",
    pages = "3611--3621",
}