MOMENT: Temporal Meta-Fact Generation and Propagation in Knowledge Graphs [F. Saïs, J. E. Gonzales Malaverri and G. Quercini @ACM-SAC-SWA 2020] This paper deals with the problem of temporal meta-fact genera- tion in RDF knowledge graphs (KGs). These temporal meta-facts represent the time validity of facts, for instance, <Barack Obama, presidentOf, United States of America> is valid for the period [2008..2016]. We propose an approach called MOMENT that com- bines two methods, the first uses a set of specified rules to generate meta-facts in knowledge bases where no temporal meta-fact exist. The second method exploits existing temporal meta-facts and a set of Horn rules generated by AMIE [9] to propagate the meta-facts and thus expand the set of temporal meta-facts. An experimental evaluation has been conducted using Yago, DBpedia and Wikidata datasets. The obtained results are promising and showed the relevance of such an approach for temporal meta-fact generation in Knowledge Graphs.

1. MOMENT – TEMPORAL META-FACT
GENERATION AND PROPAGATION IN
KNOWLEDGE GRAPHS
FATIHA SAÏS, JOANA E. GONZALES MALAVERRI,
AND GIANLUCA QUERCINI
LRI, CNRS & Paris Saclay University, France
ACM SAC-SWA, March 31h, 2020
1

2. KNOWLEDGE GRAPHS (KGS)
"Linking Open Data cloud diagram 2017, by Andrejs Abele, John
P. McCrae, Paul Buitelaar, Anja Jentzsch and Richard Cyganiak.
http://lod-cloud.net/" 2

3. facts
KNOWLEDGE GRAPHS (KGS)
"Linking Open Data cloud diagram 2017, by Andrejs Abele, John
P. McCrae, Paul Buitelaar, Anja Jentzsch and Richard Cyganiak.
http://lod-cloud.net/" 3
.
.
.

4. KNOWLEDGE GRAPHS (KGS)
4

5. 5
2012
2013
2015 2016
2013
2007
2007
2007
2012
Academic side Commercial side
WHO IS DEVELOPING
KNOWLEDGE GRAPHS?

6.  KG Creation
 Information extraction from web pages (DBpedia, Yago),
collaborative (Wikidata), ...
 KG Enrichment and Expansion
 Content prediction, data fusion, data/knowledge linking, ...
 KG Validity
 Dealing with errors and ambiguities
 Fact validity
 Timeliness: truth of statements often changes with time:
 E.g., Currently, who is the US president?
KNOWLEDGE GRAPHS: SOME
KEY PROBLEMS
6

7. KGS – WHY TIME MATTERS?
 Time period during which a fact is valid: validity time.
 f1: <#Obama presidentOf US> is true in the temporal context [2008-2016]
 f2: <#Trump presidentOf US> is true in the temporal context [2016-now]
 Query answering
 Consistency checking
 Sort the sequence of facts.
7

8. PROBLEM AND PROPOSED APPROACH
Problem
 How can we capture fact validity time using only
the facts and the structure of the KGs?
8

9. PROBLEM AND PROPOSED APPROACH
Problem
 How can we capture fact validity time using only
the facts and the structure of the KGs?
Approach – Seed meta-fact generation (step 1)
9Knowledge Patterns

10. PROBLEM AND PROPOSED APPROACH
Problem
 How can we capture fact validity time using only
the facts and the structure of the KGs?
Approach – Seed meta-fact generation (step 1)
10Knowledge Patterns

11. PROBLEM AND PROPOSED APPROACH
Temporal meta-fact
Problem
 How can we capture fact validity time using only
the facts and the structure of the KGs?
Approach – Seed meta-fact generation (step 1)
11Knowledge Patterns

12. 1. SEED META-FACT GENERATION
 Focus on predicates that may change over time:
 Brad Pitt acted in:
 the Fight Club in 1999
 the Curious Case of Benjamin Button in 2008
 Paul McCartney was/is married with:
 Heather Mills from 2002 to 2008
 Nancy Shevell since 2011
12

13. 1. SEED META-FACT GENERATION
Algorithm: Query generation
 For each knowledge pattern, in the form of:
13

14. 1. SEED META-FACT GENERATION
Algorithm: Query generation
 For each knowledge pattern, in the form of:
 Generate a SPARQL query (graph pattern = rule premise)
14

15. 1. SEED META-FACT GENERATION
Algorithm: Meta-fact generation
 On the set of facts obtained from SPARQL queries, generate the
set of temporal meta-facts
15

16. 1. SEED META-FACT GENERATION
Algorithm: Meta-fact generation
 On the set of facts obtained from SPARQL queries, generate the
set of temporal meta-facts
 For interval knowledge patterns: compute an intersection
between the temporal intervals associated to the subject and the
object of a fact.
16

17. 1. SEED META-FACT GENERATION
Algorithm: Meta-fact generation
 On the set of facts obtained from SPARQL queries, generate the
set of temporal meta-facts
 For interval knowledge patterns: compute an intersection
between the temporal intervals associated to the subject and the
object of a fact.
Start-date
End-date 17

18. TEMPORAL META-FACT
EXPANSION
18

19. 2. TEMPORAL META-FACT EXPANSION
 Use a set of Horn rules (generated by AMIE) to propagate existing
temporal meta-facts
19

20. 2. TEMPORAL META-FACT EXPANSION
 Use a set of Horn rules (generated by AMIE) to propagate existing
temporal meta-facts
 Step 1: Rule selection and instantiation
 Keep only rule with PCA-confidence > theta
 Select those premises contain a predicate in the set of seed
meta-facts
 Rule instantiation on a set of facts and seed meta-facts
(generated or already existing in the KG)
20

21. 2. TEMPORAL META-FACT EXPANSION
 Use a set of Horn rules (generated by AMIE) to propagate existing
temporal meta-facts
 Step 2: Temporal meta-fact propagation
 Case 1: only one meta-fact in the premise, then propagate the
date attached to the conclusion
 Case 2: if several meta-facts in the premise, then apply
temporal combination constraints
21

22. 2. TEMPORAL COMBINATION
CONSTRAINTS
22

23. 2. TEMPORAL COMBINATION
CONSTRAINTS
 Use of Allen’s interval relationships: during(TI, TI’), overlaps(TI, TI’),
meets(TI, TI’), …
 Apply one of the twelve combination constraints we defined, such as:
TI1 TI2
S(TI1) E(TI1)S(TI2) E(TI2)
TI3
S(TI3) E(TI3)

Inferred time interval
23

24. 2. META-FACT CONFIDENCE
Assign for each inferred meta-fact a confidence value in [0;1], using:
 Quality scores of the used rule: Head-Coverage, PCA-confidence
 K1: #meta-facts associated with the atoms of the rule premise
 K2: the sum of the distances between the temporal meta-facts of
the premise
24

25. EXPERIMENTS
25

26.  Yago: contains information extracted from
Wikipedia infoboxes, WordNet, and GeoNames,
 > 10 million entities (persons, cities,
organizations),
 > 120 million facts about these entities
 Attaches temporal and spatial dimensions to
some facts and entities.
1. SEED META-FACT GENERATION:
EXPERIMENTS ON YAGO
Time-sensitive predicates
26

27. 1. SEED META-FACT GENERATION:
EXPERIMENTS ON YAGO
Quantitative results of seed meta-fact generation
27
x 8479
x 228
x 100
x 6
x 67

28. 1. SEED META-FACT GENERATION:
EXPERIMENTS ON YAGO
Qualitative evaluation on Movie case study
- Use of IMDb as a ground truth
- Comparaison with yago meta-facts IMDb - page
28

29.  Movie domain data in YAGO:
 # of films: 151427
 # of actors: 47800
 # of release dates available: 136234
1. SEED META-FACT GENERATION:
EXPERIMENTS ON YAGO
29

30. 1. SEED META-FACT GENERATION:
EXPERIMENTS ON YAGO
 Evaluation of 5200 meta-facts of actedin predidcate
 Movie-title and actor-name equal IMDb title and one of the
IMDb actors’ name
 25% of release date in Yago are different with IMDb release
date: premiere date in IMDb and shown in a festival date in
yago
30

31. 1. SEED META-FACT GENERATION:
EXPERIMENTS ON YAGO
 Evaluation of 5200 meta-facts of actedin predidcate
 Movie-title and actor-name equal IMDb title and one of the
IMDb actors’ name
 25% of release date in Yago are different with IMDb release
date: premiere date in IMDb and shown in a festival date in
yago
 Results:
 Reach of 75% of precision for timestamp meta-facts.
 Some meta-facts are not inferred, some films have no release
date in Yago.
 Some obtained interval meta-facts are too large: need to be
refined
31

32. 2. TEMPORAL META-FACT
EXPANSION: EXPERIMENTS
 Dbpedia as a KG to be enriched with temporal meta-facts
 As seed meta-facts:
 Yago meta-facts: datasets A
 Seed meta-facts: dataset B
 Use of AMIE horn rules
 Wikidata as a ground truth (qualifiers)
 Use of Wikidata endpoint *
* https://query.wikidata.org/ 32

33. 2. TEMPORAL META-FACT
EXPANSION: EXPERIMENTS
 Quantitive results on dataset A (Yago meta-facts)
33

34. 2. TEMPORAL META-FACT
EXPANSION: EXPERIMENTS
 Quantitive results on dataset B (Seed meta-facts, step1)
34

35. 2. TEMPORAL META-FACT
EXPANSION: EXPERIMENTS
Qualitative results: Considered timestamp meta-facts and full date
 Dataset A gets better results: better accuracy of seed meta-facts
 Dataset B reaches 0.67 of F-measure
35

36. CONCLUSION
 An approach for seed temporal meta-fact
generation for time-sensitive predicates
 A rule-based approach for temporal meta-fact
propagation
 A first quantitative and qualitative evaluation
36

37. FUTURE WORK
 Develop an approach to learn time-sensitive (evolving)
predicates.
 Improvements are needed for seed interval meta-facts
generation
 Fact validity assessment using temporal-meta facts
 Combine temporal and spatial reasoning:
 E.g.: Film release dates are associated to specific
locations (country, city).
37

38. ACKNOWLEDGMENTS
 Labex DigiCosme (Paris Sacaly): VASTE
Project
 CNRS: RegleX PEPS Project
38

39. MOMENT –TEMPORAL META-FACT
GENERATION AND PROPAGATION IN
KNOWLEDGE GRAPHS
FATIHA SAÏS, JOANA E. GONZALES MALAVERRI,
AND GIANLUCA QUERCINI
LRI, CNRS & Paris Saclay University, France
ACM SAC-SWA, March 31h, 2020
39