OKGIT

Source code of the proposed models and experiments in the paper "OKGIT: Open Knowledge Graph Link Prediction with Implicit Types" to be presented in the Findings of ACL 2021.

Description

The code is based on CaRE and LAMA. The datasets used in this work can be found here. The pre-trained language models (BERT-base, BERT-large, and RoBERTa) used in the paper can be found here.

Installation

Clone the repository

git clone "https://github.com/Chandrahasd/OKGIT.git"

Install dependencies

pip install -r requirements.txt

Running

Before running the code, please edit the paths for data, config, source code, and the results in data.cfg file.

ReVerb20K

python src/main.py --dataset ReVerb20K --n_epochs 500 --model_name OKGIT --nfeats 300 --lm bert --reverse --type-loss mse --type_transform identity --type-dim 300 --type-weight 0.01 --name testrun --bmn bert-large-uncased --type_composition add --type_composition_weight 5.0 --gpu 0 --nocomet

ReVerb45K

python src/main.py --dataset ReVerb45K --n_epochs 500 --model_name OKGIT --nfeats 300 --lm bert --reverse --type-loss mse --type_transform identity --type-dim 100 --type-weight 0.0 --name testrun --bmn bert-large-uncased --type_composition add --type_composition_weight 2.0 --gpu 0 --nocomet

ReVerb20KF

python src/main.py --dataset ReVerb20KF --n_epochs 500 --model_name OKGIT --nfeats 300 --lm bert --reverse --type-loss mse --type_transform identity --type-dim 300 --type-weight 0.001 --name testrun --bmn bert-base-uncased --type_composition add --type_composition_weight 5.0 --gpu 0 --nocomet

ReVerb45KF

python src/main.py --dataset ReVerb45KF --n_epochs 500 --model_name OKGIT --nfeats 300 --lm bert --reverse --type-loss mse --type_transform identity --type-dim 300 --type-weight 0.001 --name testrun --bmn bert-base-uncased --type_composition add --type_composition_weight 0.25 --gpu 0 --nocomet

Generating Single Token Datasets

python src/preprocess/filter_triples.py --dataset <path-to-source-dataset> --bert-vocab <path-to-bert-vocab-file>

Reference

The OKGIT model is described in the following paper:

  @inproceedings{chandrahas-talukdar-2021-okgit,
  title = "{OKGIT}: {O}pen Knowledge Graph Link Prediction with Implicit Types",
  author = "Chandrahas, .  and
    Talukdar, Partha",
  booktitle = "Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021",
  month = aug,
  year = "2021",
  address = "Online",
  publisher = "Association for Computational Linguistics",
  url = "https://aclanthology.org/2021.findings-acl.225",
  doi = "10.18653/v1/2021.findings-acl.225",
  pages = "2546--2559",
  }

Chandrahasd / OKGIT