HLTCHKUST / UnifiedM2

On Unifying Misinformation Detection

Code for our "On Unifying Misinformation Detection", NAACL2021, https://arxiv.org/pdf/2104.05243.pdf

Setting Up Environment:

1. pip install -r requirement.txt to install all the dependencies.
2. From https://drive.google.com/drive/folders/1QRBKuzykREarcsn1iaNr5xKSzTGTRwu1?usp=sharing, download:
   • preprocessed.zip, unzip it and place inside ROOTDIR/misinfo/data directory. As result, you will have ROOTDIR/misinfo/data/preprocessed
   • pretrained models (unifiedM2.zip and unifiedM2_pheme.zip), unzip and place inside ROOTDIR/misinfo/pretrained directory.
     • unifiedm2.zip: pretrained UnifiedM2 model
     • unifiedM2_pheme.zip: UnifiedM2 that is trained without rumor dataset to test for PHEME's leave-one-event-out setting

Code Structure:

• /data_utils: contains codes data preprocessing and data analysis
• /model: contains codes for ST-roberta and UnifiedM2
• /utils: contains util codes for training/evaluating our models

How to run:

Refer to bash scripts listed below for your intended experiment (Before running the code, modify the arguments/parameters that fits your need/setting):

• bash run_mt.sh : for training the shared UnifiedM2 encoder

• bash run_ft.sh : for fine-tuning UnifiedM2 encoder to obtain better performance on individual tasks

• bash run_fewshot.sh : for few-shot experiment (Section 4.1)

• bash run_pheme_leave_one_event_out.sh : for leave-one-event-out experiment setting for PHEME dataset (Section 4.2)

Other details:

To add your custom tasks, made adjustments in:

• datasets.py - modify get_processor() and get_misinfo_datset() function, and add a data processor by following the format of other data processors
• utils/const.py - add your custom task name to the const file
• For other questions, send email to [e-mail] # UnifiedM2