UCL Machine Reading - FNC-1 Submission

The submission made by UCL Machine Reading (UCLMR) during Stage 1 of the Fake News Challenge (FNC-1) is based on a single, end-to-end system consisting of lexical as well as similarity features passed through a multi-layer perceptron with one hidden layer.

Although relatively simple in nature, the system performs on par with more elaborate, ensemble-based systems of other teams.

We use two simple bag-of-words representations for the text inputs: term frequency (TF) and term frequency-inverse document frequency (TF-IDF). The representations and feature thus extracted from the headline and body pairs consist of only the following:

• The TF vector of the headline;
• The TF vector of the body;
• The cosine similarity between the TF-IDF vectors of the headline and body.

A schematic overview of the setup is provided below. Further detailed information can be found in a short paper on arXiv.

Reproducibility

This repository contains the files necessary to reproduce UCLMR's submission.

Rather than providing seed values and requiring the model to be retrained, the repository contains relevant scripts and the TensorFlow model trained as part of the submission.

The submission can easily be reproduced by loading this model using the pred.py script to make the predictions on the relevant test set.

Alternatively, as suggested by the organisers of the competition, the validity of the submission can also be checked by using the pred.py script to train the model with different seeds and evaluating the mean performance of the system.

Getting started (modified)

To get started, simply download the files in this repository to a local directory.

Prerequisites

The system was developed, trained and tested using the following:

Python==3.5.2
NumPy==1.16.2
scikit-learn==0.20.3
TensorFlow==0.12.1

Please note that compatibility of the saved model with newer versions of TensorFlow has not been checked. Accordingly, please use the TensorFlow version listed above.

Setting the virtual environment

In order to avoid dependencies by the diversity of laptop setting, I would recommend to use virtual environment that already exists within the code. This is how to activate virtual environment locally.

(First, locate yourself within the repository)

• For Mac user only now (Window's instruction will be given in a day)

source bin/activate

Installing

Other than ensuring the dependencies are in place, no separate installation is required.

Simply execute the pred.py file once the repository has been saved locally.

Reproducing the submission

The pred.py script can be run in two different modes: 'load' or 'train'. Upon running the pred.py file, the user is requested to input the desired mode.

Execution of the pred.py file in 'load' mode entails the following:

• The train set will be loaded from train_stances.csv and train_bodies.csv using the corresponding FNCData class defined in util.py.
• The test set will be loaded from test_stances_unlabeled.csv and train_bodies.csv using the same FNCData class. Please note that test_stances_unlabeled.csv corresponds to the second, amended release of the file.
• The train and test sets are then respectively processed by the pipeline_train and pipeline_test functions defined in util.py.
• The TensorFlow model saved in the model directory is then loaded in place of the model definition in pred.py. The associated load_model function can be found in util.py.
• The model is then used to predict the labels on the processed test set.
• The predictions are then saved in a predictions_test.csv file in the top level of the local directory. The corresponding save_predictions function is defined in util.py. The predictions made are equivalent to those submitted during the competition.

Execution of the pred.py file in 'train' mode encompasses steps identical to those outlined above with the exception of the model being trained as opposed to loaded from file. In this case, the predictions will not be identical to those submitted during the competition.

The file name for the predictions can be changed in section '# Set file names' at the top of pred.py, if required.

Please note that the predictions are saved in chronological order with respect to the test_stances_unlabeled.csv file, however, only the predictions are saved and not combined with the Headline and Body ID fields of the source file.

Team members

• Benjamin Riedel - Full implementation
• Isabelle Augenstein - Advice
• Georgios Spithourakis - Advice
• Sebastian Riedel - Academic supervision

Citation

If you use this work in your research, please cite the short paper on arXiv using the following BibTeX entry.

@article{riedel2017fnc,
    author = {Benjamin~Riedel and Isabelle~Augenstein and Georgios~P.~Spithourakis and Sebastian~Riedel},
    title = {A simple but tough-to-beat baseline for the {Fake News Challenge} stance detection task},
    journal = {CoRR},
    volume = {abs/1707.03264},
    year = {2017},
    url = {http://arxiv.org/abs/1707.03264}
}

License

This project is licensed under the Apache 2.0 License. Please see the LICENSE.txt file for details.

Acknowledgements

• Richard Davis and Chris Proctor at the Graduate School of Education at Stanford University for the description of their development efforts for FNC-1. The system presented here is based on their setup.
• Florian Mai at the Department of Computer Science at Christian-Albrechts Universität zu Kiel for insightful and constructive discussions during system development.
• Anna Seg of FNC-1 team 'annaseg' for her suggestions on how to split the training data for more realistic system evaluation.