Automatic Detection of Entity-Manipulated Text using Factual Knowledge

Code and data used in our ACL 2022 paper titled Automatic Detection of Entity-Manipulated Text using Factual Knowledge.

Dependencies

• torch
• tensorflow
• transformers
• GPUtil
• torch_geometric
• sklearn

Data

All the data used in the paper can be downloaded from here.

Baseline Run

python roberta.py <data-path> <batch-size> <learning-rate>

where,

• <data-path>: path to the detection data (e.g., data/gpt2_123/num_entities_3)
• <batch-size>: batch size for fine-tuning (32)
• <learning-rate>: learning rate for fine-tuning (1e-5, 2e-5, 3e-5)

Ours Run

python geometric_baseline.py <data-path> <batch-size> <learning-rate> <gnn-type> <wiki-init> <gnn-num-feats> <mp-iter> <first-hop-neighbors-path> <entity-supervision>

where,

• <data-path>: path to the detection data (e.g., data/gpt2_123/num_entities_3)
• <batch-size>: batch size for fine-tuning (32)
• <learning-rate>: learning rate for fine-tuning (1e-5, 2e-5, 3e-5)
• <gnn-type>: type of GNN (e.g., GCNConv)
• <wiki-init>: need to initialize node embeddings with wikipedia2vec embeddings? (0 or 1) (ensure WIKI_EMB_F in the code is set to path of the wikipedia2vec embeddings)
• <gnn-num-feats>: hidden dimension of the GNN (300)
• <mp-iter>: number of message passing iterations (aka depth of GNN) (1, 2, 3)
• <first-hop-neighbors-path>: path to the first hop neighbors of each node (e.g., data/kb_first_hop_neighbors_entity_replace/nonlm_123_leastfreq/num_entities_3)
• <entity-supervision>: need entity supervision? (0 or 1)

Cite

If you find this project useful, please cite the paper:

@inproceedings{jawahar-etal-2022-automatic,
    title = "Automatic Detection of Entity-Manipulated Text using Factual Knowledge",
    author = "Jawahar, Ganesh  and
      Abdul-Mageed, Muhammad  and
      Lakshmanan, Laks",
    booktitle = "Association for Computational Linguistics",
    year = "2022",
    pages = "86--93"
}

Contact

If you have any questions or suggestions, please contact Ganesh Jawahar.

License

This repository is GPL-licensed.