Knowledge Representation Learning with Entities Attributes and Relations

Knowledge Representation Learning with Entities, Attributes and Relations Yankai Lin, Zhiyuan Liu, Maosong Sun Tsinghua University May 22, 2016

Knowledge Graph Completion • Goal – Infer the missing fact in Knowledge Graph 2

Knowledge Graph Completion • Based on Graph Features • Based on Representation Learning – Tensor Factorization – Implicit Space – Geometric space 3

Based on Graph Features • Path Ranking Algorithm – Path Features – Classifier • Example – school 4

Tensor Factorization • RESCAL 5

Implicit Space • • Structure Embedding (SE) Semantic Matching Energy (SME) Neural Tensor Network (NTN) … 6

Geometric space • Trans. E – Embedding: • Entity Vectors • Relation Vectors – Goal： • Extension – Trans. H, Tran. R, Trans. D … 7

Can we do better? • Fact in Knowledge Graph – Entity Attributes – Relations between Entities • Embed with different Models! 8

Existing Problem • Attributes vs. Relations 9

Our Model • Relational Triple Encoder • Attributional Triple Encoder 10

Relational Triple Encoder • Traditional Knowledge Representation Model – Trans. E – Trans. R –… 11

Attributional Triple Encoder • Classifier – Neural Network –… • Attribute Correlations – Profession • Gender • Education –… 12

Empirical Evaluation • Data – FB 24 k • Task – Entity Prediction – Relation Prediction – Attributes Prediction 13

Entity Prediction • (head, relation, ? ) • (? , relation, tail) 14

Relation Prediction • Correlation between Relations & Attributes (CRA) – General Form of Type Constraint 15

Attribute Prediction • Attribute Correlations (AC) 16

Example 17

Conclusion • Distinguish existing KG-relations into attributes and relations • Propose a new KR model (KR-EAR) • Encode the correlations between entity attributes in KR-EAR. 18

Future Work • Combine Probabilistic Graphic Model with Knowledge Representation • Split relations and attributes by held-out methods 19