Keras '2.0.8'

Hyperopt '0.1'

1. Download th data from competition Codalab.

https://competitions.codalab.org/competitions/19938

2. Download 300d word vectors pre-trained on Google News corpus.

https://code.google.com/archive/p/word2vec/

3. Change filepaths for data and for word embeddings if needed:

in help_prep_functions.py in loadW2vModel() function insert filepath for word embeddings

in preprocessing_tweets.py and preprocessing_reddit.py change filepaths for data if needed.

4. Choose features option:

In prep_pipeline.py on line 98:

def main(data ='RumEval2019', feats = 'SemEvalfeatures')

feats can be either text for avgw2v representation of the tweets or SemEvalfeatures for additional extra features concatenated with avgw2v.

5. Download necessary nltk packages, if they have not been downloaded.

In Python interactive interpreter, input the following commands:

import nltk

nltk.download('punkt')
nltk.download('averaged_perceptron_tagger')

6. Run preprocessing script

python prep_pipeline.py

The description of the model architecture can be found in https://www.aclweb.org/anthology/S/S17/S17-2083.pdf The features used in this code are different to the ones used in the paper.

1. In outer_semeval2019.py you can choose the number of trials that the search algorithm performs while searching for the parameter combination.

2. In parameter_search.py you can define search_space.

3. Make sure output/ folder has been created

mkdir output

4. Run the baseline

python outer_semeval2019.py

If you have any questions feel free to contact me [email protected] or other task organisers [email protected]