industry-eval-EA

The code and benchmark of paper An Industry Evaluation of Embedding-based Entity Alignment [arxiv] [coling] in Proceedings of COLING 2020.

Code

We present the source code to generate biased seed mappings for EA.

code
|__ check_benchmark.py
|__ sample_benchmark.py
|__ config.json

Specifically, we present a total of four settings in extracting biased seed mappings:

• baseline [without any bias]
  • "Ideal" in 4.2 of our paper
  • "With No Bias" in 4.3 of our paper
• name-biased [same name]
  • "With Name Bias" in 4.3 of our paper
• attr-biased [more attributes]
  • "With Attribute Bias" in 4.3 of our paper
• industry [same name & more attributes]
  • "Industrial" in 4.2 of our paper

all of which follow the algorithm introduced in 3.2 of our paper.

To check the validity of any to-be-used benchmark, please run check_benchmark.py to verify the benchmark format.

To generate/sample biased seed mappings from to-be-used benchmark, please run sample_benchmark.py and later check train/val/test splits in the target_root_dir defined in config.json.

To reproduce the experimental results in our paper, please refer to OpenEA to run the experiments based on the biased seed mappings (mentioned above).

examples

We list some statistics of sampled biased seed mappings as follows.

<D_W_15K_V2> train_ratio: 0.02 val_ratio: 0.01
train-name-bias-stats:   same1.00  close0.00  diff0.00
train-attr-bias-stats:   large1.00  mid0.00  small0.00
val-name-bias-stats:   same1.00  close0.00  diff0.00
val-attr-bias-stats:   large1.00  mid0.00  small0.00
test-name-bias-stats:   same0.32  close0.21  diff0.47
test-attr-bias-stats:   large0.57  mid0.32  small0.10

When sampling biased seed mappings from the public benchmark D_W_15K_V2 under the industry setting (both name-biased and attribute-biased), it can be seen that train/val splits only contains biased seed mappings, which have the same name and large number of attributes.

Benchmark

We extracted the industry benchmark, named MED-BBK-9K, from two real-world medical KGs for alignment, which can be found here.

industry.zip
|__ attr_triples_1          # attribute triples of KG1
|__ attr_triples_2          # attribute triples of KG2
|__ ent_links               # entity links between KGs (ground-truth)
|__ rel_triples_1           # relation triples of KG1
|__ rel_triples_2           # relation triples of KG2

The statistics of the industry benchmark is listed as follows. D_W_15K_V2 is also recorded for the purpose of comparison.

Benchmark	KGs	#Ents	#Rels	#Rel triples	Rel degree	#Attrs	#Attr triples	Attr degree
MED-BBK-9K	MED	9,162	32	158,357	34.04	19	11,467	1.24
MED-BBK-9K	BBK	9,162	20	50,307	10.96	21	44,987	4.91
D-W-15K	DBpedia	15,000	167	73,983	8.55	175	66,813	4.40
D-W-15K	Wikidata	15,000	121	83,365	10.31	457	175,686	11.59

examples

We list some fragments of our industry benchmark as follows.

ent_links

<月经异常>\t<月经不调>
<弓形体病性巩膜炎>\t<弓形虫病性巩膜炎>
<巨趾症>\t<巨趾症>
<发细菌感染>\t<a40292>
<脑溢血后遗症>\t<脑出血后遗症>

rel_triples

<骨关节病>\t<典型症状>\t<僵硬>
<额区感觉减退>\t<相关疾病>\t<下肢动脉硬化闭塞症>
<绦虫病>\t<典型症状>\t<恶心>
<胆汁返流性胃炎>t<典型症状>\t<反酸>
<脐炎>\t<典型症状>\t<发热>

attr_triples

<病毒性食管炎>\t<英文名>\t<viralesophagitis>
<碱中毒>\t<临床表现>\t<它是呼吸系统对碱中毒的代偿现象，借助于浅而慢的呼吸，得以增加肺泡内的pco，使[bhco] [hhco]的分母加大，以减少因分子变大而发生的比值改变（稳定ph值）。躁动、兴奋、谵语、嗜睡、严重时昏迷。有手足搐搦，腱反射亢进等。如已发生钾缺乏，可能出现酸性尿的矛盾现象，应特别注意。标准碳酸氢（sb）、实际碳酸氢（ab）、缓冲碱（bb）、碱剩余（be）增加，血液paco、血液ph值升高。>
<十二指肠溃疡>\t<就诊科室>\t<消化内科>

Citation

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{zhang2020industry,
  title={An Industry Evaluation of Embedding-based Entity Alignment},
  author={Zhang, Ziheng and Liu, Hualuo and Chen, Jiaoyan and Chen, Xi and Liu, Bo and Xiang, YueJia and Zheng, Yefeng},
  booktitle={Proceedings of the 28th International Conference on Computational Linguistics: Industry Track},
  pages={179--189},
  year={2020}
}