This repository contains code used in our paper:
"ACTC: Active Threshold Calibration for Cold-Start Knowledge Graph Completion"
to be presented at ACL'23 🚀
by Anastasiia Sedova and Benjamin Roth.
For any questions please get in touch.
ACTC is a new method for estimation the relation threshold for a cold-start knowledge graph completion. ACTC leverages a limited set of labeled and a large set of unlabeled data in order to calculate per-relation thresholds. Basing on these thresholds and plausibility scores calculated by a knowledge graph embedding model, one can make a decision about whether a new triple should be included to the knowledge graph or not. Mostly important, it helps to find thresholds in a setting where there is only a limited set of available manual annotations.
The ACTC could be launched by running the main.py script.
Here is an example:
python main.py
--path_to_data path/to/directory/with/data/and/KGE/model/predictions/
--output_dir path/to/output/directory
--path_to_config path/to/config/file
An example of a directory where data and KGE model predictions are stored (for CoDEx-s dataset + ComplEx embeddings): data
An example of a config file: scripts/configs/config.json
When using our work, please cite our ArXiV preprint:
@inproceedings{sedova-roth-2023-actc,
title = "{ACTC}: Active Threshold Calibration for Cold-Start Knowledge Graph Completion",
author = "Sedova, Anastasiia and Roth, Benjamin",
booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers)",
month = jul,
year = "2023",
address = "Toronto, Canada",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2023.acl-short.158",
pages = "1853--1863"
}
This research has been funded by the Vienna Science and Technology Fund (WWTF)[10.47379/VRG19008] and by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) RO 5127/2-1.