Completeness, Recall, and Negation
in Open-World Knowledge Bases
1. Introduction & foundations (Simon)
2. Predictive recall assessment (Fabian)
3. Counts from text and KB (Shrestha)
4. Negation (Hiba)
5. Relative completensss & wrap‐up (Simon)
Simon Razniewski, Hiba Arnaout, Shrestha Ghosh, Fabian Suchanek
CC-BY
Fabian M. Suchanek
. To get a PDF of these slides, open in Chrome and print to PDF.
>ruleMining
Predictive recall assessment
How can we find out if a knowledge base is complete?
•
The Basics: Predicting facts
• Recall of facts
•
Do we have all objects for a subject?
•
Can we use text to determine completeness?
• Recall of entities
•
Do we have all entities of the real world?
2
Fact Prediction Problem
3
[Roslyn Pinker from
Pinterest
, Steven Pinker by
Rose Lincoln
, Susan Pinker by
De Standaard
]
hasChild
hasChild
gender
female
Steven Pinker
Susan Pinker
Roslyn Pinker
Fact Prediction Problem
4
Problem:
Fact Prediction Problem
Input:
a knowledge base Ｋ
Task:
Find facts ｆ∕∈Ｋ that are true in the real world.
We may be able to deduce some facts that are very likely to be true
in reality, even though they are not in the KB.
=> This is where the KB must be incomplete
hasSister
hasChild
hasChild
gender
female
Fact Prediction by Rule Mining
5
Given a KB, rule mining automatically finds logical rules such as:
... usually with a confidence score. These can be used to predict facts.
ｈａｓＣｈｉｌｄ（ｘ，ｙ） ∧ ｈａｓＣｈｉｌｄ（ｘ，ｚ） ∧ ｇｅｎｄｅｒ（ｚ， ｆｅｍａｌｅ） ⇒ ｈａｓＳｉｓｔｅｒ（ｙ，ｚ）
ｍａｒｒｉｅｄＴｏ（ｘ，ｙ） ∧ ｈａｓＣｈｉｌｄ（ｘ，ｚ） ⇒ ｈａｓＣｈｉｌｄ（ｙ，ｚ）
ｗａｓＢｏｒｎＩｎ（ｘ，ｙ） ∧ ｈａｓＬａｎｇｕａｇｅ（ｘ，ｚ） ⇒ ｓｐｅａｋｓ（ｘ， ｚ）
hasSister
hasChild
hasChild
gender
female
>details
Fact Prediction by Rule Mining
6
Bottom‐up approaches
Start with rules for concrete instances, generalize them
Top‐down approaches
Start with short rules, make them longer
Jonathan Lajus, Luis Galárraga, Fabian M. Suchanek:
“
Fast and Exact Rule Mining with AMIE 3
”, ESWC 2020
Stefano Ortona, Venkata Vamsikrishna Meduri, Paolo Papotti:
“
Robust Discovery of Positive and Negative Rules in Knowledge Bases
”
(Rudik system) ICDE 2018
C. Meilicke, M. Chekol, D. Ruffinelli, H. Stuckenschmidt:
“
Anytime Bottom-Up Rule Learning for Knowledge Graph Completion
”
(AnyBurl system), IJCAI 2019
>linkPred
Fact Prediction by Link Prediction
7
We can try to embed the entities in an ｎ ‐dimensional vector space
in such a way that their relative position corresponds to their relations: