Commonsense Knowledge Base Completion with Structural and Semantic Context
Code for the paper Commonsense Knowledge Base Completion with Structural and Semantic Context.
Bibtex
@article{malaviya2020commonsense,
title={Commonsense Knowledge Base Completion with Structural and Semantic Context},
author={Malaviya, Chaitanya and Bhagavatula, Chandra and Bosselut, Antoine and Choi, Yejin},
journal={Proceedings of the 34th AAAI Conference on Artificial Intelligence},
year={2020}
}
Requirements
- PyTorch
- Run
pip install -r requirements.txtto install the required packages.
Dataset
The ATOMIC dataset used in this paper is available here and the ConceptNet graph is available here. For convenience, both the pre-processed version of ATOMIC and ConceptNet used in the experiments are provided at this link.
Note: The ATOMIC dataset was pre-processed to canonicalize person references and remove punctuations (described in preprocess_atomic.py.
Note: The original evaluation sets provided in the ConceptNet dataset contain correct as well as incorrect tuples for evaluating binary classification accuracy. valid.txt in data/conceptnet is the concatenation of the correct tuples from the two development sets provided in the original dataset while test.txt is the set of correct tuples from the original test set.
Training
To train a model, run the following command:
python -u src/run_kbc_subgraph.py --dataset conceptnet --evaluate-every 10 --n-layers 2 --graph-batch-size 60000 --sim_relations --bert_concat
This trains the model and saves the model under the saved_models directory.
Language Model Fine-tuning
In this work, we use representations from a BERT model fine-tuned to the language of the nodes in the knowledge graph.
The script to fine-tune BERT as a language model on the two knowledge graphs is present in the lm_finetuning/ directory. For example, here is a command to fine-tune BERT as a language model on ConceptNet:
python lm_finetuning/simple_lm_finetuning.py --train_corpus {CONCEPTNET_TRAIN_CORPUS} --bert_model bert-large-uncased --output_dir {OUTPUT_DIR}
Pre-Trained Models
We provide the fine-tuned BERT models and pre-computed BERT embeddings for both ConceptNet and ATOMIC at this link. If you unzip the downloaded file in the root directory of the repository, the training script will load the embeddings.
We also provide the pre-trained KB completion models for both datasets for ease of use. Link to Conceptnet model and ATOMIC model.
Evaluation
To evaluate a trained model, and get predictions, provide the model path to the --load_model argument and use the --eval_only argument. For example, to evaluate the pre-trained ConceptNet model provided above, use the following command:
CUDA_VISIBLE_DEVICES={GPU_ID} python src/run_kbc_subgraph.py --dataset conceptnet --sim_relations --bert_concat --use_bias --load_model {PATH_TO_PRETRAINED_MODEL} --eval_only --write_results
This will load the pre-trained model, and evaluate it on the validation and test set. The predictions are saved to ./topk_results.json.
Similarly, to evaluate the trained model on ATOMIC, use the following command:
CUDA_VISIBLE_DEVICES={GPU_ID} python src/run_kbc_subgraph.py --dataset atomic --sim_relations --use_bias --load_model {PATH_TO_PRETRAINED_MODEL} --eval_only --write_results
Please email me at [email protected] for any questions or comments.
152 Dec 27, 2022
60 Dec 29, 2022
60 Dec 31, 2022
23 Dec 16, 2022
61 Dec 30, 2022
130 Dec 13, 2022
33 Dec 5, 2022
35 Sep 8, 2021
37 Dec 4, 2022
0 Oct 17, 2022
80 Nov 28, 2022
3 Nov 30, 2021
73 Nov 6, 2022
146 Nov 29, 2022
41 Nov 30, 2022
296 Dec 6, 2022
68 Dec 28, 2022
19 Nov 30, 2022
24 Oct 13, 2022