Source code of paper: "Online Opinion Conflict Interaction Recognition Based on Dependent Multi-Task Deep Learning".
Accepted by ICWSM 2026 (The 20th International AAAI Conference on Web and Social Media, Los Angeles).
This is the online appendix https://github.com/zss019/CIMDL/blob/main/online_appendix.pdf for the paper, which includes “Appendix A. Prompts for LLM in the Experiment” and “Appendix B. Case Analysis”. These appendices constitute the supplementary material for the experimental analysis in the paper.
ScienceNet_dataset and zhihu_dataset are two benchmark datasets we have constructed, which have been divided into train/dev/test sets. We crawled users' opinion interaction records and manually annotated them.
Zhihu dataset consists of comments and corresponding replies in Q&A thread on selected controversial topics. These interactions were annotated as “Support” and “Conflict”. ScienceNet dataset includes comments and replies on the web-blogs about scientific or academic topics. In addition to “Support” and “Conflict” , we have also annotated “Neutral” for some smooth interaction.
-
Python 3.10
-
PyTorch 2.1.0
-
RTX 3090 GPU
-
CUDA 12.1
Recommended to use GPU for training (such as Google Colab, AutoDL, Aliyun, etc.)
The CIMDL.py file located in the models folder is the model file, which contains the definition scheme of the model.
The run.py, train_eval.py, and utils.py files are responsible for the project's operation, training, and processing respectively.
Additionally, the pre-trained model we used was downloaded from: https://huggingface.co/hfl/chinese-roberta-wwm-ext and https://huggingface.co/DMetaSoul/sbert-chinese-qmc-domain-v1.
Part of code (particularly data processing and loading) references the [Multi-architecture Open-source Chinese Text Classification] (https://github.com/649453932/Chinese-Text-Classification-Pytorch) project. Readers interested in our paper can refer to that project's documentation for details..
python run.py --model CIMDL