The code and benchmark of paper OntoEA: Ontology-guided Entity Alignment via Joint Knowledge Graph Embedding [arxiv][acl] in Findings of ACL-IJCNLP 2021.
The source code of OntoEA is implemented based on OpenEA and we follow the same design features of OpenEA to achieve functionality while maintaining extensibility. We retain the minimal fraction of OpenEA in the implementation of OntoEA, and we list the source code (along with comments) of as follows.
code
|__ run
| |__ args
| | |__ ontoea_args_15K.json // OntoEA config file for 15K benchmark
| | |__ ontoea_args_100K.json // OntoEA config file for 100K benchmark
| |__ main_from_args.py // OntoEA interface with args input
| |__ run_OntoEA.sh // bash script to run experiments of OntoEA
|__ src
| |__ openea
| | |__ approaches
| | | |__ ontoea.py // OntoEA core code
| | |__ models
| | | |__ basic_model.py // same as OpenEA
| | |__ modules
| | | |__ args
| | | | |__ args_hander.py // same as OpenEA
| | | |__ base
| | | | |__ initializers.py // same as OpenEA
| | | | |__ losses.py // same as OpenEA
| | | | |__ mapping.py // same as OpenEA (with minor updates)
| | | | |__ optimizers.py // same as OpenEA
| | | |__ finding
| | | | |__ alignment.py // same as OpenEA
| | | | |__ evaluation.py // same as OpenEA (with minor updates)
| | | | |__ similarity.py // same as OpenEA
| | | |__ load
| | | | |__ kg.py // same as OpenEA
| | | | |__ kgs.py // same as OpenEA (with updates)
| | | | |__ read.py // same as OpenEA (with updates)
| | | |__ train
| | | | |__ batch.py // same as OpenEA (with minor updates)
| | | | |__ sample.py // same as OpenEA (with minor updates)
| | | |__ utils
| | | | |__ check.py // same as OpenEA (with updates)
| | | | |__ inference.py // same as OpenEA (with updates)
| | | | |__ util.py // same as OpenEA
Since we adopt the OpenEA codebase to implement OntoEA, the way to run OntoEA is much similar to the way to use OpenEA. We provide two ways to run OntoEA:
1. using bash scripts (recommended)
Change into ./code/run and run the bash script
./run_OntoEA.sh
2. using off-the-shelf approaches with python interface
Change into ./code/run and run OntoEA as
python main_from_args.py <predefined_arguments> <benchmark_name> <dataset_splits>
and for example
python main_from_args.py ./args/ontoea_args_15K.json MED_BBK_9K 721_5fold/1/
Different settings of OntoEA can be configured in the ontoea_args_15K.json and ontoea_args_100K.json. The key configurations are listed as follows.
{
"training_data": "../../benchmarks/", // where the benchmark is
"output": "../../output/results/", // where the output results go
"dataset_division": "721_5fold/1/", // where the train/valid/test splits are
"word_embed": "../../benchmarks/wiki-news-300d-1M.vec", // where the word embeddings are
"use_word_embed_init": 0, // whether or not to use word embeddings in OntoEA (i.e., 1 == OntoEA w/ SI; 0 == OntoEA w/o SI)
"use_word_embed_init_onto": 0, // whether or not to use word embeddings in OntoEA (i.e., 1 == OntoEA w/ SI; 0 == OntoEA w/o SI)
"word_embed_init_zh": 0, // whether or not Chinese word embeddings (only for MED_BBK_9K)
"use_alter_label": 0, // whether or not to use alternative labels for entity names (only for D_W benchmarks)
}
Some other configurations on the hyperparameters (lambda_1, lambda_2, lambda_3) that balance different losses are listed as follows.
{
"alpha": 5, // controlling L_E (entity embedding)
"gamma": 1, // controlling L_M (membership loss)
"sigma": 1, // controlling L_C (ontology embedding)
}
Therefore, the experiments of Model Component Analysis in Sec. 4.3 of our paper can be re-produced using these settings:
w/o L_C ==> to set sigma = 0
w/o L_M ==> to set gamma = 0
w/o Onto. ==> to set sigma = 0 and gamma = 0
- Python 3.x (tested on Python 3.6.12)
- Tensorflow 1.x (tested on Tensorflow 1.14.0)
- Scipy
- Numpy
- Pandas
- Scikit-learn
- Gensim
All the dependencies can be installed using the provided requirements.
The newly constructed benchmarks extends the original KG alignment benchmarks (OpenEA and MED_BBK_9K) with ontological information, including ontology triples (both disjointWith and subClassOf), membership links (or cross-view links), ontology hierarchy.
The data files are identical for each benchmark, and we list all the benchmarks with MED_BBK_9K benchmark details as follows.
benchmarks
|__ EN_FR_15K_V1 // EN_FR_15K_V1 from OpenEA with appended ontological information
|__ EN_FR_15K_V2 // EN_FR_15K_V2 from OpenEA with appended ontological information
|__ EN_DE_15K_V1 // EN_DE_15K_V1 from OpenEA with appended ontological information
|__ EN_DE_15K_V2 // EN_DE_15K_V2 from OpenEA with appended ontological information
|__ D_W_15K_V1_A // D_W_15K_V1 from OpenEA with appended ontological information and ontologies pre-aligned with PARIS
|__ D_W_15K_V1_M // D_W_15K_V1 from OpenEA with appended ontological information and ontologies pre-aligned with manual annotation
|__ D_W_15K_V2_A // D_W_15K_V2 from OpenEA with appended ontological information and ontologies pre-aligned with PARIS
|__ D_W_15K_V2_M // D_W_15K_V2 from OpenEA with appended ontological information and ontologies pre-aligned with manual annotation
|__ MED_BBK_9K // MED_BBK_9K with appended ontological information
| |__ 721_5fold/1 // train/valid/test splits
| |__ attr_triples_1 // attribute triples of KG1
| |__ attr_triples_2 // attribute triples of KG2
| |__ rel_triples_1 // relation triples of KG1
| |__ rel_triples_2 // relation triples of KG2
| |__ ent_links // all entity mappings between KG1 and KG2
| |__ crossview_link_1 // membership links between KG1 and its ontology
| |__ crossview_link_2 // membership links between KG2 and its ontology
| |__ onto_attr_triples // attribute triples of the aligned/shared ontology
| |__ onto_disjointWith_triples // disjointWith relation triples of the aligned/shared ontology
| |__ onto_subClassOf_triples // subClassOf relation triples of the aligned/shared ontology
| |__ class_path.json // class hierarchy of the aligned/shared ontology
Note that D_W_15K_V1/V2_A are constructed with the ontologies pre-aligned with alignment system PARIS and D_W_15K_V1/V2_M are constructed with the ontologies pre-aligned with manual annotation. These two different settings correspond to the Analysis of Ontology Alignment Methods in Sec. 4.3 of our paper.
If you have any difficulty or question in running code and reproducing experimental results, please email to zihengzhang1025@gmail.com
If you use this model or code, please cite it as follows:
@inproceedings{xiang-etal-2021-ontoea,
title = "{O}nto{EA}: Ontology-guided Entity Alignment via Joint Knowledge Graph Embedding",
author = "Xiang, Yuejia and
Zhang, Ziheng and
Chen, Jiaoyan and
Chen, Xi and
Lin, Zhenxi and
Zheng, Yefeng",
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.96",
doi = "10.18653/v1/2021.findings-acl.96",
pages = "1117--1128",
}