Practical Applications of Temporal and Event Reasoning

1
Practical Applications of Temporal and Event
Reasoning

• James Pustejovsky, Brandeis
• Graham Katz, Osnabrück
• Rob Gaizauskas, Sheffield
• ESSLLI 2003
• Vienna, Austria
• August 25-29, 2003

2
Course Outline

• Monday-
• Theoretical and Computational Motivations
• Overview of Annotation Task
• Events and Temporal Expressions
• Tuesday
• Anchoring Events to Times
• Relations between Events
• Wednesday
• Syntax of TimeML Tags
• Semantic Interpretations of TimeML
• Relating Annotations
• Temporal Closure
• Thursday
• Automatic Identification of Expressions
• Automatic Link Construction
• Friday-
• Outstanding Problems

3
Friday Topics

• Events with Argument Binding
• TimeML German Fragment
• Outstanding Problems
• TimeML-enabled Applications

4
Features in TimeML 2.0

• Argument binding into Events
• Pred feature in EVENT
• General types with like entailments
• Vendler classification
• Scope of Negation and Modality
• Represented on TLINK

5
Argument Binding into Events
6
Syntax of Entity

• ltENTITYgt
• attributes aid type agreement det
• aid ID
• aid argumentID
• argumentID altintegergt
• type ltnamed entity typegt
• agreement ???
• det athepossessivequant

7
Syntax of Arglink

• ltARGLINKgt
• attributes lid origin eventInstanceID
  (relatedEventInstanceID relatedArgumentID)
  preposition
• lid ID
• lid LinkID
• LinkID lltintegergt
• origin CDATA
• eventInstanceID IDREF
• eventInstanceID EventInstanceID
• relatedEventInstanceID IDREF
• eventInstanceID EventInstanceID
• relatedArgumentID IDREF
• argumentID argumentID
• preposition CDATA

8
Example of Arguments 1

• John left on Saturday.
• ltENTITY aida1gt John lt/ENTITYgt
• ltEVENT eide1 classOCCURRENCEgt
• left
• lt/EVENTgt
• ltMAKEINSTANCE eiidei1 eide1 tensepast
  aspectnone/gt
• ltSIGNAL sids1gt on lt/SIGNALgt
• ltTIMEX3 tidt1 value20011014T112713gt Saturday
  lt/TIMEXgt
• ltARGLINK eventInstanceIDei1 relatedEntityIDa1
  /gt
• ltTLINK eventInstanceIdei1 relatedToTimet1
  RelType IS_INCLUDED/gt

9
Example of Arguments

• Police arrested the suspect in the airport on
  Tuesday.
• ltENTITY aida1gt Police lt/ENTITYgt
• ltEVENT eide1 classOCCURRENCEgt
• arrested
• lt/EVENTgt
• ltMAKEINSTANCE eiidei1 eide1 tensepast
  aspectnone/gt
• ltENTITY aida2 detthegt the suspect lt/ENTITYgt
• in
• ltENTITY aida3 detthegt the airport lt/ENTITYgt
• ltSIGNALgt on lt/SIGNALgt
• ltTIMEX3 tidt1 value20011014T112713gt Saturday
  lt/TIMEXgt
• ltARGLINK eventInstanceIDei1 relatedEntityIDa1
  /gt
• ltARGLINK eventInstanceIDei1 relatedEntityIDa2
  /gt
• ltARGLINK eventInstanceIDei1 relatedEntityIDa3
  prepostionin/gt
• ltTLINK eventInstanceIdei1 relatedToTimet1
  RelType IS_INCLUDED/gt

10
Negation over Events Currently

• Survivors
• were
• not
• ltEVENT eid"e1" class"OCCURRENCE"
  pred"TEACH"gt
• found
• lt/EVENTgt
• ltSIGNAL sid"s1"gt
• on
• lt/SIGNALgt
• ltTIMEX3 tid"t2" type"DATE" value"XXXX-WXX-1"gt
• Monday
• lt/TIMEX3gt
• ltMAKEINSTANCE eventID"e1" eventInstanceID"ei1"
  " negationTRUE"gt
• ltTLINK eventInstanceID"ei1" signalID"s1"
  relatedToTime"t2" relType"IS_INCLUDED"/gt
• No survivors were found.
• ltMAKEINSTANCE eventID"e1" eventInstanceID"ei1"
  " negationFASLE"gt
• ltTLINK eventInstanceID"ei1" signalID"s1"
  relatedToTime"t2" relType"IS_INCLUDED"/gt

11
Quantifiers and Negation 1

• Survivors were not found on Monday.
• ltENTITY aida1gt Survivors lt/ENTITYgt
• Were
• ltSIGNAL gt not lt/SIGNALgt
• ltEVENT eide1 classOCCURRENCE predfind
  tensepast aspectnonegt
• found
• lt/EVENTgt
• ltSIGNAL sid s1gt on lt/SIGNALgt
• ltTIMEX3 tidt1 value20011014T112713gt Monday
  lt/TIMEXgt
• ltARGLINK eventIDei1 relatedEntityIDa1 /gt
• ltTLINK eventIdei1 relatedToTimet1
  Polarity NEG RelType IS_INCLUDED/gt

12
Quantifiers and Negation 2

• No survivors were found on Monday.
• ltENTITY aida1 quantNOgt No survivors
  lt/ENTITYgt
• were
• ltEVENT eide1 classOCCURRENCE predfind
  tensepast aspectnonegt
• found
• lt/EVENTgt
• ltSIGNAL sid s1gt on lt/SIGNALgt
• ltTIMEX3 tidt1 value20011014T112713gt Monday
  lt/TIMEXgt
• INTENDED INTERPRETATION
• ltARGLINK eventIDei1 relatedEntityIDa1 /gt
• ltTLINK eventIdei1 relatedToTimet1
  Polarity FALSE RelType IS_INCLUDED/gt
• Reference to the Argument (no survivors)
  provides a resource to the interpretation
  function for determining the polarity of the
  TLINK.

13
TimeML German Fragment

• (Due to Frank Schilder, ms. 2003)

14
TimeML in German

• Corpus study in German focussing on the
  preposition in.
• Ca. 100 occurrences of the preposition in
  extracted from taz articles
• Marked with simplified TimeML
• Only TLINKS
• Different Aspect specification
• Marked with additional features (see below)
• Goal definition of a semantics for the
  proposition in considering
• Aspectual classes
• Granularity
• Reference time

Schilder (2003)
15
German temporal and event expressions

• Different tense and aspect system
• Usage of tenses
• Present tense is ambiguous wrt. Present/future
  tense.
• No progressive form
• (Past perfect preferred tense in spoken language
  for expressing past events)
• Aspectual information not morphologically encoded
  in a consistent way
• Different Aktionsarten
• Ingressive verlieben (to fall in love)
• Exgressive verblühen (to wither)
• Semelfactive husten (to cough)
• Iterative hüsteln (coughing)
• No imperfective/perfective aspect

Schilder (2003)
16
German temporal and event expressions

• Different syntactical structure
• Prefix-verbs ausschließen (exclude) / schließen
  (close) Die Bedingungen schließen einen Verkauf
  aus
• Reflexive verbs sich entwickeln (come out) /
  entwickeln (develop)
• Complex verb constructions sich in der Lage
  sehen etwas zu tun (feel capable of doing
  something)
• Sah sich die Polizei schon bisher nicht in der
  Lage
• , dass die die Polizei sich schon bisher nicht in
  der Lage sah
• Normally, Verbs are at position 2, but
• Participle verbs come at the end of a clause and
• Subordinate clauses and relative clauses end of
  clause

Schilder (2003)
17

• Schröder hatte bereits am Wochenende
  signalisiert, dass er eine dritte Amtszeit
  anstrebt.
• 29.8.03
• ltENTITY aida1gt Schröder lt/ENTITYgt
• hatte
• bereits
• ltSIGNALgt am lt/SIGNALgt
• ltTIMEX3 tidt1 type date valueYYYY-XX-WEgt
  Wochenende lt/TIMEXgt
• ltEVENT eide1 classOCCURRENCEgt signalisiert
  lt/EVENTgt
• ltMAKEINSTANCE eiidei1 eide1 tensepast
  aspectnone/gt
• ltSLINK eventInstanceID"ei1" subordinatedEvent"ei
  2 relType"MODAL"/gt
• ltTLINK eventInstanceIdei1 relatedToTimet1
  RelType IS_INCLUDED/gt
• dass
• ltENTITY aida2 gt er lt/ENTITYgt
• ltENTITY aida3 deteinegt eine dritte Amtszeit
  lt/ENTITYgt
• ltEVENT eide1 classOCCURRENCEgt anstrebt
  lt/EVENTgt
• ltMAKEINSTANCE eiidei2 eide2 tensepresent
  aspectnone/gt
• ltARGLINK eventInstanceIDei1 relatedEntityIDa1
  /gt
• ltARGLINK eventInstanceIDei1 relatedEventIDei2
  /gt

18
Outstanding Problems
19
Semantics of TimeML

• A text T is satisfied by a model M iff there are
  functions
• fe Dome -gt Pow(E), and
• fei Domei -gt E
• ft Domt -gt I , such that
• for all tags t ?Tag(T), t is satisfied by fe fei
  and ft in M.
• A tag t is satisfied by fe,ft, and fei in M iff
  if t has the form
• ltEVENT eid ? class ? pred ? gt then fe(?)
  Val(?)
• ltTIMEX3 tid ? type DATE value ? gt then
  ft(?) Val(?)
• ltTLINK eventInstanceID ? relatedtoTime ?
  relType IS_INCLUDEDgt then ?(fei(?)) ? ft (
  ?)

20
Problems for Interpretation

• Negation
• John didnt teach on TuesdayltEVENT eid
  ei1predTEACHgt ltMAKEINSTANCE eiid
  eii1negationtruegtltTIMEX3 tidt1
  valXXXX-WXX-2gtltTLINK relatedToTime t1
  eventInstanceeii1 relationIS_INCLUDEDgt
• -gt SCOPE for negation
• Multiple Events
• John taught twice on TuesdayltEVENT eid
  ei1predTEACHgt ltMAKEINSTANCE eiid eii1
  evente1 cardinality2gtltTIMEX3 tidt1
  valXXXX-WXX-2gtltTLINK relatedToTime t1
  eventInstanceeii1 relationIS_INCLUDEDgt
• ltMAKEINSTANCE eiid ? eid ? negationFALSE
  modal cardinality?gt then fei(?) ? fe(?)
  ltMAKEINSTANCE eiid eii1 evente1
  gtltMAKEINSTANCE eiid eii2 evente1 gt
• Condition on Embedding Functions

21
Problems for TimeML

• Set-valued Times
• John taught ltTIMEX3 tidt4 typeSET
  valueP1M quantEVERY freq3Dgt three days
  every monthlt/TIMEX3gt
• PROBLEM the temporal identifier cant be
  interpreted as denoting a particular interval of
  time, it must be a set of intervals (or even a
  set of sets of intervals!)
• Disjunction
• John taught on Monday or on Wednesday

22
Some Solutions

• Negation
• Use TLINK as a scope domain, eliminate
  MAKEINSTANCE
• John didnt teach on TuesdayltEVENT eid
  ei1predTEACHgt ltTIMEX3 tidt1
  valXXXX-WXX-2gtltTLINK relatedToTime t1
  eventIDei1 relationIS_INCLUDED
  Polarityfalsegt
• New TLINK Rules
• ltTLINK eventID ? relatedtoTime ? relType
  IS_INCLUDED Polaritytruegt there is an e ?
  E such that e ? fe(?) and ?(e) ? ft ( ?)
• ltTLINK eventID ? relatedtoTime ? relType
  IS_INCLUDED Polarityfalsegt
• there is no e ? E such that e ? fe(?) and ?(e) ?
  ft ( ?)

23
Some Solutions

• Multiple events
• Add cardinality element to the TLINK
• John taught twice on TuesdayltEVENT eid
  ei1predTEACHgt ltTIMEX3 tidt1
  valXXXX-WXX-2gtltTLINK relatedToTime t1
  eventIDei1 relationIS_INCLUDED
  Polaritytrue cardinality2gt
• ltTLINK eventID ? relatedtoTime ? relType
  IS_INCLUDED Polaritytrue cardinality?gt is
  satisfied iffthere are Val(?) distinct e ? E
  such that e ? fe(?) and ?(e) ? ft ( ?)

24
Harder Problems

• Vagueness
• When he left, shortly after 5 am Tuesday, he
  discovered someone had smashed a window.
• Appavu has been involved with healthcare
  standards development for about a decade, an
  interest he developed shortly after he began
  working with information systems at Cook County.
• Domino's Pizza of Washington reported that they
  delivered "In excess" of 100 large pizzas to the
  White House late this afternoon.
• It was then,early in December of 1977, that he
  went to the NORML conference.

25
Vagueness

• Current Treatment
• ltTIMEX3 tidt1 val20030829TAFT modENDgt
  late this afternoon lt/TIMEX3gt
• ltTIMEX3 tidt1 val197712 modSTARTgt early
  in December of 1977 lt/TIMEX3gt
• Problem
• No semantics for mod attributes means no
  possibility for doing reasoning.
• It was then,early in December of 1977, that he
  went to the NORML conference. Two weeks later he
  was a convert.
• Before or after Christmas?
• We might fake a solution by being overly general
• Interpret START to mean the first half of

26
Current Treatment

• No general solution for mod values
• Shortly after 5am -gt minutes
• Shortly after he began working -gt weeks or months

27
Semantic Weakness

• Simple annotation of temporal relations is too
  week
• President John F. Kennedy's gravesite at
  Arlington National Cemetery has been restored to
  its original condition, after someone tried
  unsuccessfully to dig up some of its granite
  paving stones.
• South Africa, after losing the toss, were bowled
  out for 107 against England.
• How long after?
• Days or weeks
• An hour or two.
• This is not generally encoded overtly.

28
Context Dependent Vagueness

• If we did code this, lots of world-knowledge
  based information could be encoded by annotators
• They ate lunch early on Monday.
• They ate dinner early on Monday.
• They ate breakfast early on Monday.
• Probably
• before noon
• in the early evening
• in the very early morning

29
Questions

• How to talk about a likely distribution in
  time?
• How to compare such annotations?

30
TimeML-enabled Applications
31
Web-basedTemporal Reasoning

• Web Negotiation Agents (Brokers)
• Scheduling Programs

32
Semantic Web
Delivery within five business days. ltEVENT
eid"e1" class"OCCURRENCE" tense"NONE"
aspect"NONE"gt order lt/EVENTgt ltMAKEINSTANCE
eiid"ei1" eventID"e1"/gt ltEVENT eid"e2"
class"OCCURRENCE" tensenil aspect"PERFECTIVE"gt
delivery lt/EVENTgt ltMAKEINSTANCE eiid"ei2"
eventID"e2"/gt ltSIGNAL sid"s1"gt within lt/SIGNALgt
ltTIMEX3 tid"t1" type"DURATION" valuenil
temporalFunction"true"gt five business
days lt/TIMEX3gt ltTLINK eventInstanceID"ei2"
signalID"s1" relatedToEvent"ei1
relType"AFTER" /gt ltTLINK eventInstanceID"ei2"
signalID"s1" relatedToTime"t1
relType"IS_INCLUDED" /gt ltTLINK
eventInstanceID"ei1" signalID"s1"
relatedToTime"t1 relType" BEGINS" /gt
33
Scheduling Issues

• Mary teaches on Mondays and Wednesdays in the
  fall.
• Sophie goes to daycare on Thursday and Friday at
  400pm in October.

34
www.TimeML.org