Coupled Hierarchical Transformer

Preprocessed SemEval-2017 and PHEME Datasets and Codes for our paper: Coupled Hierarchical Transformer for Stance-Aware Rumor Verification in Social Media Conversations (https://aclanthology.org/2020.emnlp-main.108.pdf).

Author

Jianfei Yu

jfyu@njust.edu.cn

July 15, 2020

Data (All the datasets are under the "rumor_data" folder)

• Stance Classification (SemEval-2017): the "semeval17_stance" folder.
• Rumor Verification: SemEval-2017 (the "semeval17" folder) & PHEME (the "pheme" folder)

Requirement

• PyTorch 1.0.0
• Python 3.7

Code Usage

Preprocessing

• Step 1: Process the stance classification dataset (SemEval-2017)

python process_stance_semeval17.py

• Step 2: Process the rumor verification dataset (SemEval-2017 & PHEME)

python process_rumor_semeval17.py

python process_rumor_pheme.py

Model Details

• Hierarchical Transformer for Stance Classification: each conversation thread is split into 10 subthreads, and each subthread is set to have 17 tweets, and each tweet has 30 tokens.
• Hierarchical Transformer and Coupled Hierarchical Transformer for Rumor Verification: Due to memory limitation, each conversation thread is split into 4 subthreads, and each subthread is set to have 17 tweets, and each tweet has 30 tokens.
• Note that for each subthread, you can change the number of tweets and the number of words in each tweet by tuning the two parameters: --max_tweet_num 17 --max_tweet_length 30. Also, it is required that max_tweet_num * max_tweet_length <= 512.

Training for Stance Classification with Hierarchical Transformer (Single-task Learning)

• SemEval-2017: This is the training code of tuning parameters on the dev set, and testing on the test set. Note that you can change "CUDA_VISIBLE_DEVICES" based on your available GPUs.

sh run_multi_stance_semeval17_10BERT.sh

Training for Rumor Verification with Hierarchical Transformer (Single-task Learning)

• SemEval-2017: This is the training code of tuning parameters on the dev set, and testing on the test set. Note that you can change "CUDA_VISIBLE_DEVICES" based on your available GPUs.

sh run_rumor_semeval17.sh

• PHEME: This is the training code of performing 9 fold cross validation. Note that you can change "CUDA_VISIBLE_DEVICES" based on your available GPUs.

sh run_rumor_pheme.sh

Training for Rumor Verification with Coupled Hierarchical Transformer (Multi-task Learning)

• SemEval-2017: This is the training code of tuning parameters on the dev set, and testing on the test set. Note that you can change "CUDA_VISIBLE_DEVICES" based on your available GPUs.

sh run_multitask_rumor_stance.sh

• PHEME: This is the training code of performing 9 fold cross validation. Note that you can change "CUDA_VISIBLE_DEVICES" based on your available GPUs.

sh run_multitask_rumor_stance_pheme.sh

We show our running logs on the SemEval-2017 dataset in the folder "running_log_files", and you can find similar results as reported in our EMNLP submission.

Acknowledgements

• Using these two datasets means you have read and accepted the copyrights set by Twitter and dataset providers.