Dataset

[lujiaying]

~- [ ] DBLP~ ~- [ ] PubMed~ ~- [ ] Wiki-subset~

[lujiaying]

DBLP

• 156K paper abstracts version: HiExpan, CoRel
• 79K paper abstracts version: NetTaxo
• 4M paper abstracts: DBLP-Citation-network V12

[lujiaying]

What relations can be generalized?

<Monkey, eat, banana>

[lujiaying]

Setting on May 17

• close world link prediction:
• open world link prediction: first induce the taxonomy (isA ?), then induce relation.
• taxonomy relations and non-taxonomy relations to predict under the ranking setting

Two datasets

• WN18RR
• ConceptNet Subset or CN-100K

[lujiaying]

Dataset Stat

Dataset	#nodes	#edges	#taxo_edges	#non-taxo_edges	setting
WN18RR	40K, 5K, 5K	86K, 3K, 3K	3.7K, 1.2K, 1.3K	4.9K, 1.7K, 1.7K	close-world, 96.1% test ent in train
CN-100K	78K, 1.8K, 1.3K	100K, 1.2K, 1.2K	15K, 197, 211	85K, 1003, 989	close-world, 92.8% test ent in train

Preliminary Results

	WN18RR MRR	H@1	H@3	H@10	CN100K MRR	H@1	H@3	H@10
TransE	.226			.501
DistMult	.43	.39	.44	.49	.09	.05	.10	. 17
ConvE	.43	.40	.44	.52	.21	.14	.23	.34
ComplEx	.44	.41	.46	.51	.11	.07	.12	.19
RotateE	.476	.428	.492	.571
HAKE	.497	.452	.516	.582
Most-Freq	.026	.015	.025	.045	.073	.035	.060	.144
Hard-Taxo					.072	.033	.060	.147
CBR	.43	.39	.46	.51
Prob-CBR	.48	.43	.49	.57
Taxo-CBR

• Hard-Taxo: test_h_has_sib=627 (0.522), test_t_has_sib=416 (0.347)

WN18RR from HAKE CN-100K from Commonsense Knowledge Base Completion with Structural and Semantic Context

[lujiaying]

Heuristic Baselines

Commonsense Knowledge Base Completion Similar Fact Count, Argument Similarity, Max Similarity: these train binary classification models.

Most-Frequent: <h, r, ?> then ? is the most freq tail ent for r.

Hard-Taxonomy

• infer taxo:
  • <h, IsA, ?>: find similar h' (score func), then <h', IsA, t'> gives t' scores to rank (propagate from h).
  • <?, IsA, t>: h similar to existing <h', IsA, t>.
• infer non-taxo: try to induce taxo, then generalize over taxo.
  • <h, r, ?>: suggested by sibling and similar items
  • <?, r, t>: suggested by sibling and similar items

[lujiaying]

Case Based Reasoning A Simple Approach to Case-Based Reasoning in Knowledge Bases, https://openreview.net/forum?id=AEY9tRqlU7

Replace the k-nearest search into taxonomy-guided similar entity search.

Case study

WN18RR

to predict: 06845599 (__trade_name) _member_of_domain_usage 03754979(__metharbital)

similar ents: {'06333653' (name): 1, '06338908' (designation): 1, '06336537' (street_name): 1, '06338278' (pseudonym): 1, '06334512' (fictitious_name): 1, '06343520' (title): 1, '06337307' (__given_name): 1, '06335532' (__filename): 1}

retrieved fact: {('06336537'(__street_name), '_member_of_domain_usage', '02837416' (__thunder, street names for heroin)): 1.0})

CN100K

triple to predict: dog CapableOf bark ranked cor_hs: [('dog', 'CapableOf', 'live'), ('dog', 'CapableOf', 'have sex'), ('dog', 'CapableOf', 'lie down'), ('dog', 'CapableOf', 'sleep'), ('dog', 'CapableOf', 'hear'), ('dog', 'CapableOf', 'relaxation'), ('dog', 'CapableOf', 'have fun'), ('dog', 'CapableOf', 'move'), ('dog', 'CapableOf', 'relax'), ('dog', 'CapableOf', 'listen to music'), ('dog', 'CapableOf', 'shelter'), ('dog', 'CapableOf', 'object'), ('dog', 'CapableOf', 'build'), ('dog', 'CapableOf', 'poop'), ('dog', 'CapableOf', 'live in')] MRR=0.002 hits@1,3,10 =0.000, 0.000, 0.000 triple to predict: pen UsedFor write ranked cor_hs: [('pen', 'UsedFor', 'learn'), ('pen', 'UsedFor', 'live in'), ('pen', 'UsedFor', 'education'), ('pen', 'UsedFor', 'learn thing'), ('pen', 'UsedFor', 'study'), ('pen', 'UsedFor', 'shelter'), ('pen', 'UsedFor', 'store thing'), ('pen', 'UsedFor', 'teach student'), ('pen', 'UsedFor', 'teach'), ('pen', 'UsedFor', 'listen to teacher'), ('pen', 'UsedFor', 'store object'), ('pen', 'UsedFor', 'child'), ('pen', 'UsedFor', 'play game'), ('pen', 'UsedFor', 'teach child'), ('pen', 'UsedFor', 'make friend')] MRR=0.011 hits@1,3,10 =0.000, 0.000, 0.000