Multitask4Veracity

This repository contains code for the paper "All-in-one: Multi-task Learning for Rumour Stance classification,Detection and Verification" by E. Kochkina, M. Liakata, A. Zubiaga

This code relies on preprocessed data that can be downloaded at https://figshare.com/articles/PHEME_dataset_Preprocessed_for_Multitask_Learning_for_Rumour_Verification/6473873

Raw data can be downloaded at https://figshare.com/articles/PHEME_dataset_for_Rumour_Detection_and_Veracity_Classification/6392078

How to run the code:

Install prerequisites

• Python 3
• Keras
• Hyperopt
• Optparse

Run outer.py

outer.py has the following options:

• --model - which task to train, stance or veracity
• --data - which dataset to use
• --search - boolean, controls whether parameter search should be performed
• --ntrials - if --search is True then this controls how many different parameter combinations should be assessed
• --params - specifies filepath to file with parameters if --search is false
• -h, --help - explains the command line

running

python outer.py

will be equivalent to running:

python outer.py --model='mtl2stance' --data='RumEval' --search=True --ntrials=10 --params="output/bestparams.txt" 

MTL2 Veracity + Stance

RumEval

python outer.py --model='mtl2stance' --data='RumEval' --search=True --ntrials=50

FullPHEME

5 folds

python outer.py --model='mtl2stance' --data='PHEME5' --search=True --ntrials=50

or

9 folds

python outer.py --model='mtl2stance' --data='PHEME9' --search=True --ntrials=50

MTL2 Veracity + Detection

5 folds

python outer.py --model='mtl2detect' --data='PHEME5' --search=True --ntrials=50

or

9 folds

python outer.py --model='mtl2detect' --data='PHEME9' --search=True --ntrials=50

MTL3 Veracity + Stance + Detection

5 folds

python outer.py --model='mtl3' --data='PHEME5' --search=True --ntrials=50

or

9 folds

python outer.py --model='mtl3' --data='PHEME9' --search=True --ntrials=50