1. Knowledge Base Completion
周天烁
2017年9月6日

2. Outline Introduction Approach Outlook Latent feature model
Word Embedding
Observed feature model
Path Ranking Algorithm
Markov logic networks
Outlook
Highlight work
Frontier research

3. Introduction Knowledge graph arises Knowledge base completion
Datasets prosper
Suffer from incompleteness/ heterogeneity/ ambiguity
Knowledge base completion/ instances matching/ entity resolution
Knowledge base completion
triple classification /link prediction (h,?,t)

4. observed feature model
Approach
KBC
latent feature model
Markov logic networks
observed feature model

5. Approach ——latent feature model
General idea
entity representation：vector/tensor
relation : operations on the entity’s vector space
typical models:
TransE
RESCAL
MLP

6. Approach ——latent feature model
TransE—translation on word embedding
from word2vec to word embedding

7. Approach ——latent feature model
TransE—translation on word embedding
translation on word embedding
V(king) - V(man) + V(woman) ≈ V(queue)
h+r=?t
(head,relation,tail)

8. Approach ——latent feature model
other models
fijk=eJiWkej

9. Approach ——latent feature model
pros
powerful
Scalability
cons
low interpretability
dependency on hyperparameter

10. Approach ——observed feature model
intuition
similar entities are likely to be related (homophily)
the similarity of entities can be derived from the paths between nodes
typical models:
Inductive Logic Programming (ILP)
Path Ranking Algorithm（PRA）

11. Approach ——observed feature model
Path Ranking Algorithm（PRA)

12. Approach ——observed feature model
Path Ranking Algorithm（PRA)

13. Approach ——observed feature model
Path Ranking Algorithm（PRA)

14. Approach ——observed feature model
Path Ranking Algorithm（PRA)
pros
high interpretability
improvement space
cons
Scalability
efficiency

15. Approach ——markov logic network
Marckov
Random field
first order logic
markov logic network

16. Approach ——markov logic network
pros
suitable for good structure;
e.g. NLP(chain),CV(grid); GIS(coordinate）；
cons
computational intractable
rule leaning

17. Outlook highlight frontier research related academic area
word embedding
PRA
rule leaning
frontier research
hybrid model
DeepLearning for word embeddings
PRA refinement
common sense knowledge reasoning
search personalization
related academic area
statistical relational learning（SRL）
link prediction in other networks

18. Reference
[Dat Quoc Nguyen,2017] An overview of embedding models of entities and relationships for knowledge base completion. Yang, Fan, Zhilin Yang, and William W. Cohen. "Differentiable Learning of Logical Rules for Knowledge Base Completion." arXiv preprint arXiv: (2017). Lin, Xixun, Yanchun Liang, and Renchu Guan. "Compositional learning of relation paths embedding for knowledge base completion." arXiv preprint arXiv: (2016). [Bordes et al., 2011] Learning structured embeddings of knowledge bases [Bordes et al., 2013] Translating embeddings for modeling multi-relational data. [Nickel et al., 2015] A review of relational machine learning for knowledge graphs. [Lao and Cohen, 2010] Relational retrieval using a path-constrained random walks. [Lao et al., 2011] Random walk inference and learning in a large scale knowledge base. M. Richardson and P. Domingos, “Markov logic networks,” Machine Learning, vol. 62, no. 1, pp. 107–136, Wang, Quan, Bin Wang, and Li Guo. "Knowledge Base Completion Using Embeddings and Rules." IJCAI Gardner, Matt, and Tom M. Mitchell. "Efficient and Expressive Knowledge Base Completion Using Subgraph Feature Extraction." EMNLP zhouzhihua. Machine Learning c5eb1484a8582f&utm_source=tuicool&utm_medium=referral

19. Q & A
Thanks for listening