CoDEx: A Comprehensive Knowledge Graph Completion Benchmark

EMNLP 2020  ·  Tara Safavi, Danai Koutra ·

We present CoDEx, a set of knowledge graph completion datasets extracted from Wikidata and Wikipedia that improve upon existing knowledge graph completion benchmarks in scope and level of difficulty. In terms of scope, CoDEx comprises three knowledge graphs varying in size and structure, multilingual descriptions of entities and relations, and tens of thousands of hard negative triples that are plausible but verified to be false. To characterize CoDEx, we contribute thorough empirical analyses and benchmarking experiments. First, we analyze each CoDEx dataset in terms of logical relation patterns. Next, we report baseline link prediction and triple classification results on CoDEx for five extensively tuned embedding models. Finally, we differentiate CoDEx from the popular FB15K-237 knowledge graph completion dataset by showing that CoDEx covers more diverse and interpretable content, and is a more difficult link prediction benchmark. Data, code, and pretrained models are available at https://bit.ly/2EPbrJs.

PDF Abstract EMNLP 2020 PDF EMNLP 2020 Abstract

Datasets


Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Link Prediction CoDEx Large TuckER MRR 0.309 # 2
Hits@1 0.244 # 2
Hits@3 0.3395 # 2
Hits@10 0.430 # 2
Link Prediction CoDEx Large RESCAL MRR 0.304 # 3
Hits@1 0.242 # 3
Hits@3 0.3313 # 3
Hits@10 0.419 # 4
Link Prediction CoDEx Large ConvE MRR 0.303 # 4
Hits@1 0.240 # 4
Hits@3 0.3298 # 4
Hits@10 0.420 # 3
Link Prediction CoDEx Large ComplEx MRR 0.294 # 5
Hits@1 0.237 # 5
Hits@3 0.3179 # 5
Hits@10 0.400 # 5
Link Prediction CoDEx Large TransE MRR 0.187 # 6
Hits@1 0.116 # 6
Hits@3 0.2188 # 6
Hits@10 0.317 # 6
Link Prediction CoDEx Medium TransE MRR 0.303 # 6
Hits@1 0.259 # 3
Hits@3 0.3599 # 3
Hits@10 0.458 # 4
Link Prediction CoDEx Medium ComplEx MRR 0.337 # 2
Hits@1 0.244 # 4
Hits@3 0.3477 # 5
Hits@10 0.456 # 5
Link Prediction CoDEx Medium RESCAL MRR 0.317 # 5
Hits@1 0.239 # 5
Hits@3 0.3551 # 4
Hits@10 0.464 # 3
Link Prediction CoDEx Medium TuckER MRR 0.328 # 3
Hits@1 0.223 # 6
Hits@3 0.3363 # 6
Hits@10 0.454 # 6
Link Prediction CoDEx Medium ConvE MRR 0.318 # 4
Hits@1 0.262 # 2
Hits@3 0.3701 # 2
Hits@10 0.476 # 2
Link Prediction CoDEx Small TuckER MRR 0.444 # 2
Hits@1 0.372 # 2
Hits@3 0.5038 # 2
Hits@10 0.646 # 2
Link Prediction CoDEx Small ComplEx MRR 0.404 # 4
Hits@1 0.293 # 5
Hits@3 0.4494 # 5
Hits@10 0.623 # 6
Link Prediction CoDEx Small TransE MRR 0.354 # 6
Hits@1 0.339 # 4
Hits@3 0.4975 # 3
Hits@10 0.638 # 3
Link Prediction CoDEx Small RESCAL MRR 0.404 # 4
Hits@1 0.343 # 3
Hits@3 0.4926 # 4
Hits@10 0.635 # 4
Link Prediction CoDEx Small ConvE MRR 0.444 # 2
Hits@1 0.219 # 6
Hits@3 0.4218 # 6
Hits@10 0.634 # 5

Methods


No methods listed for this paper. Add relevant methods here