HERO: Linguistic-style-aware Neural Networks for Fake News Detection

The implementation of HERO in the paper: Linguistic-style-aware Neural Networks for Fake News Detection

Require

PyTorch >= 1.9.1

nltk >= 3.6.3

About input data

1. We use the dataset ReCOVery as an example in the folder ReCOVery which has been split as training set, validation set and test set. For more detail about the ReCOVery dataset, we can refer to the webstite ReCOVery.

2. We use Stanford's GloVe 100d word embeddings as word embedding in this paper, which is named as glove.6B.100d.txt in our code. The file of word embeddings can be downloaded from the webstite, Glove.6b.100d.

3. For processing RST and CFT of ReCOVery dataset (or other news dataset), we use the code from the following website Generate RST and CFG tree.

4. Here we provide a simple example of the format of RST and CFG in the folder /data/strtree_RST and /data/strtree_CFG. We generate RST and CFG tree for the example news Original_text_news_1.txt, which finally produces news_1.txt (RST) and news_1.txt (CFG) in the folders /data/strtree_RST and /data/strtree_CFG respectively.

Reproducing Results

1. When getting all the input data including words embedding file glove.6B.100d.txt, folder /data/strtree_RST and /data/strtree_CFG, we can use the following commnads with the trained model in the result folder to reproduce the result on the ReCOVery dataset.

        python test.py