1. Temporal Constraint Acquisition
Joint work with Partha Talukdar and Tom Mitchell

2. Motivation Automatically built Knowledge Bases (KB) (e.g., NELL, Yago)
Contain millions of facts, but miss critical dimension: time
Temporal orders temporal scopes
fact: actedIn (Angelina Jolie, The Tourist) in 2010
constraint: actedIn (Person, Film) is during directedBy (Film, Person)
inference: directedBy (The Tourist, Florian Henckel) in 2010
Collective Temporal Scoping, CoTS (Talukdar et al., 2012)
Hand-coded temporal constraints among facts
ILP-based joint optimization for temporal scoping
constrain
Can we automatically induce these temporal constraints?
e.g., actedIn (Person, Film) happens during directedBy (Film, Person)

3. Problem Statement
Automatically induce temporal relationships (before, during, etc) between relations
Relations are groups of verbs
E.g. actedIn(Person, Film):
Angelina Jolie acted in The Tourist
Pirates of the Caribbean starring Johnny Depp

4. Comparison with Previous Approach
(Schank and Abelson, 1977; Mani et al., 2006; Chambers and Jurafsky, 2008)
Our approach
Hand-coded / supervised
Self supervised
Document specific
(Micro-reading)
Over a large number of documents
(Macro-reading)
At the level of events = verbs
At the level of relations = group of verbs

5. Key Ideas Relation mentions are expressed by verbs
Narrative sequence  temporal sequence
Irregularities are ironed out when aggregated over many documents

6. Approach

7. Find Verbs that Express Relations

8. Find Verbs that Express Relations
actedIn (Person, Film)
actedIn (Angelina Jolie, The Tourist)
SVO triples
Subject
Object
Angelina Jolie starred in The Tourist
Verb
890 million dependency-parsed sentences
from ClueWeb09 dataset

9. Find Occurrences of Verbs

10. Find Occurrences of Verbs
starring
directed
nominated
actedIn (Person, Film)
directedBy (Film, Person)
hasWonAward (Film, Award)
during
before

11. Infer Temporal Order from Narrative Order

12. Infer Temporal Order from Narrative Order
Pairwise Temporal Ordering of Relations r1 r2 r3 v1 v3 v5
RelVerb(r1, v1)
RelVerb(r2, v3)
RelVerb(r3, v5)
NBef(v1, v3)
NDur(v2, v5)
TBefore(r1, r2)
TDuring(r2, r3)

13. Infer Temporal Order from Narrative Order
Pairwise Temporal Ordering of Relations r1 r2 r3 v1 v3 v5
RelVerb(r1, v1)
RelVerb(r2, v3)
RelVerb(r3, v5)
NBef(v1, v3)
NDur(v2, v5)
TBefore(r1, r2)
TDuring(r2, r3)

14. Infer Temporal Order from Narrative Order
Collective Temporal Ordering of Relations
using graph-based semi supervised learning algorithm r1 r2 r3 v1 v3 v5
RelVerb(r1, v1)
RelVerb(r2, v3)
RelVerb(r3, v5)
NBef(v1, v3)
NDur(v2, v5)
TBefore(r1, r2)
TDuring(r2, r3)
TBefore(r1, r3)

15. Experimental Setup Relations from Yago2 KB to time-order
Relation instances to time-scope
Protagonists: albums, cricketers, films, footballers, novels, operas, plays, politicians, and songs

16. Temporal Ordering Experiment
Compute DAG of learned temporal orderings
Hand-construct gold standard ordering
Compare to random baseline

17. Temporal Ordering Experiment

18. Temporal Ordering Experiment
Politicians
Films

19. Temporal Scoping Experiment
CoTS to time scope 1831 relation instances
69.8% precision in the inferred scope previously unknown in Yago2
Examples of correct scoping:
actedIn(Jason Biggs, American Pie 2) is in 2001
actedIn(John Neville, High School High) is in 1996

20. Conclusion
Automatic induction of temporal ordering constraints between relations
Based on narrative order of verbs expressing these relations, aggregated over many documents
Observe effectiveness for the problem of temporal ordering and temporal scoping