NTMGraph

The data and code of paper Topic Aware Graph.

1. Preprocess the data Gossipcop

The format of dataset of the model NTM we used is in ntm/data/gossipcop_lines.txt, which you can find that each line in the txt file is one doc of your datas.

So first to preprocess to get the txt file:

$ python preprocess.py

And in the preprocess.py you can change the data path.

The dataset gossipcop_v3_keep_data_in_proper_length.json we offered here has 15,729 news.

Just clik the link gossipcop_v3_keep_data_in_proper_length.json to download it.

The 0-3,784 are fake news, and the other 3785-1,5729 are real news. Detailes can be seen in CossipCop-LLM.

2. Run the NTM model

$ cd ntm
$ python GSM_run.py --taskname gossipcop

Then you will find the model ckpt in ckpt/, and change the ckpt path in the next code embedding.py.

$ python embedding.py

After that you could get the docs embedding under topic in results/gossipcop_embed.py, whose size is (n, 16).(The n_topic we set is 16).

The docs of different topic is stored in results/gossipcop_topic_{i}.txt, where i is from 0 to 15.

3. Construct Topic Aware Graph

$ cd ..
$ python graph_construct.py

Then you can get different 16 graphs of 16 topics in data/sparse_matrix_topic_{j}.npz, where j is from 0 to 15.

4. Training

$ python main.py

After that you will see the scores print on the terminal, and you can see the epoch-accuracy graph on the tensorboard.

More optional arguments can also be seen in main.py.

Reference

This work has received assistance from the following. Consider citing their works if you find this repo useful.

@misc{ZLL2020,
  author = {Leilan Zhang},
  title = {Neural Topic Models},
  year = {2020},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/zll17/Neural_Topic_Models}},
  commit = {f02e8f876449fc3ebffc66f7635a59281b08c1eb}
}

https://github.com/SZULLM/GossipCop-LLM