The Encoding of Lexical Implications in VerbNet

1
The Encoding of Lexical Implications in VerbNet

• Change of Location Predicates
• Annie Zaenen, Danny Bobrow and Cleo Condoravdi

2
Outline

• Rationale of the talk and aims of the project
• VerbNet information
• Conclusions

3
Reusability

• Lexical resources are expensive to create
• The best way to create them is collaboratively
  and to structure them in such a way that they can
  be used for several different projects
• In how far is VerbNet information reusable?

4
Inferences about locations context

• Application Question Answering/textual inference
• Method
• create a Knowledge Representation for a text,
  e.g. a newspaper article and a question/inference
• subsumption calculation to see whether the
  information in the question is contained in the
  text
• Approach normalize the text to a logical form
  through rewrite rules.

5
Example 1
6
Aim of subproject

• Use VN information to make inferences about
  change of location of event participants
• E.g.
• Annie went from San Francisco to Morocco
• gt
• Annie was in San Francisco at t1 Annie was in
  Morocco at t2 t1 lt t2
• In our representation

7
Representation before change of location
calculation

• Annie left San Francisco for Morocco.
• Conceptual Structure
• subconcept(leave7,leave-1,)
• role(Theme,leave7,Annie1)
• role(Source,leave7,San Francisco12)
• role(Destination,leave7,Morocco30)
• subconcept(Annie1,female-2)
• subconcept(Morocco30,location-1,location-4)
• subconcept(San Francisco12,location-1,location-
  4)
• Temporal Structure
• temporalRel(startsAfterEndingOf,Now,leave7)
• within square brackets information from WN

8
Target Representation 2

• Annie left San Francisco for Morocco.
• Conceptual Structure
• subconcept(locate48,locate-1,locate-3)
• subconcept(locate47,locate-1,locate-3)
• role(Theme,locate48,Annie1)
• role(Theme,locate47,Annie1)
• role(Location,locate48,Morocco30)
• role(Location,locate47,San Francisco12)
• subconcept(leave7,leave-1)
• role(Source,leave7,San Francisco12)
• role(Theme,leave7,Annie1)
• role(Destination,leave7,Morocco30)
• subconcept(Annie1,female-2)
• subconcept(Morocco30,location-1,location-4)
• subconcept(San Francisco12,location-1,location-4
  )
• Temporal Structure

9
Getting the necessary information

• How do we know that the moving object in a
  sentence like John left New York is John?
• Where do we get information that tells us that
  before the event described in the sentence John
  is in New York and that afterwards he is no
  longer there?

10
Outline

• Rationale and aims of the project.
• VerbNet information
• VerbNet Semantics
• Change of Location predicates in VerbNet
• Conclusions

11
Can we use VerbNet information for this?

• Levin classes
• VN2-1 239 XML-files representing Levin
  (sub)classes some additions

12
VerbNet

• Verb classes in VN are based on Levin
  classification. This classification embodies the
  belief that there is a close correspondence
  between (some aspects of) the meaning of verbs
  and their subcategorization alternations (John
  sank the boat the boat sank).
• Verbs are classified based on their
  subcategorization alternations.
• VerbNet adds the thematic role information and a
  semantic representation.

13
VN information

• Class Send-11.1
• Thematic roles
• Agent, Theme, Source, Destination
• Selectional restrictions
• Agentanimate ororganization,
• Themeconcrete,
• Sourcelocation,
• Destinationlocation
• Frames
• Name NP-PP-PP
• Example Nora sent the book from London to Paris.
  
• Syntax Agent V Theme Source Destination
• Semantics
• cause(agent,E)
• motion(during(E),Theme)
• location(start(E),Theme, Source)
• location(end(E),Theme,Destination)

14
Why not use VN semantic roles?

• Theme used for participants in a location or
  undergoing a change of location.
• Agent generally a human or animate subject.
  Used mostly as a volitional agent, but also used
  in VerbNet for internally controlled subjects
  such as forces and machines.
• Destination end point of the motion, or
  direction towards which the motion is directed.
• Source start point of the motion. Usually
  introduced by a source prepositional phrase
  (mostly headed by from or out of.)
• Location underspecified destination, source or
  place in general introduced by a locative or path
  prepositional phrase.
• from Kipper (2005)

15
Use VN semantics

• Event structure in VN based on Moens and
  Steedman (1988)
• Not all events have culminations if there is no
  culmination there is no result state.
• John built a house culmination (accomplishment)
• John played the piano no culmination (activity)

16
VN semantics (event structure)

• Send Agent, Theme, Destination
• Amanda sent the package to New York.
• motion(during(E), Theme),
• location(end(E), Theme, Destination),
• cause(Agent, E0)

17
VN semantics (event structure)

• Shove ltPREP value"to towards"gt
• motion(during(E), Theme),
• location(end(E), Theme, Destination),
• cause(Agent, E)

18
A problem that we will ignore

• No free variation to/towards
• Prepositions (Kipper Schuler, 2005) Spatial
• path
• src from, out, out of,
• dir across, along, around, down,
• dest
• dest-conf into, onto,
• dest-dir for, at, to, towards,
• loc about, above, against,
• What is
• Further distinction necessary
• They sent the kids into the mountains.
• They slid the books onto the table.

19
Combinations

• Slide-11.2 The books slid from the desk to the
  floor.
• motion(during(E), Theme),
• location(start(E), Theme, Source),
• location(end(E), Theme, Destination)
• Carry-11.4 Nora carried the books to Paris.
• equal(E0,E1)
• motion(during(E0), Theme),
• location(end(E0), Theme, Destination),
• motion(during(E1), Agent)
• location(end(E0), Agent, Destination),
• cause(Agent, E)
• Send-11.1 Nora sent the books to London.
• motion(during(E), Theme),
• location(end(E), Theme, Destination),
• cause(Agent, E)

20
Another promising pattern

• Class-9.1 put
• motion(during(E), Theme),
• not(Prep(start(E), Theme, Destination)),
• Prep(end(E), Theme, Destination),
• cause(Agent, E)
• Class-9.3.2 The water rushed into the house.
• motion(during(E), Theme),
• not(Prep(start(E), Theme, Destination)),
• Prep(end(E), Theme, Destination)
• Here the idea is that the value of the
  preposition has to be factored in. How the VN
  people saw this exactly is not of our concern.
  Given the right semantics for the preposition we
  can use the information.

21
Small classes

• class 16-concealment
• (She hid the presents in the drawer.),
  location(result(E),Patient, Location)
• class-22 (mix, shake, tape, etc.)
• mingled(result(E),physical,Patient1,Patient2)
  together(end(E),physical,Patient1,Patient2)
• class-23
• together(start(E),physical,Patient1,Patient2),
  apart(end(E),physical,Patient1,Patient2)
• class-47.5.2
• not(together(start(E),physical,Themei ,Themej
  )), together(end(E),physical,Themei ,Themej )
• The representation is not analytic enough to get
  an invariant representation of change of location
  in these cases

22
Incomplete coverage in classes that are covered
in principle

• class 9.3 (funnel), only endpoints are given
• Funnel the liquid from the bottle into the cup.
• class 9.5 (pour), no frame with both start and
  end points
• He poured the water from the bowl into the cup
• class 9.7
• OK Jessica loaded boxes into the wagon
• Not Jessica loaded the boxes from the train into
  the car.
• 10.2 (banish) and class 10.4.2 (shovel) both a
  source and a destination frame are given but no
  frame that combines the two.
• Shovel the snow from the sidewalk into the ditch.

23
Potential VN classes with start and end locations

• Classes that seemed good candidates to me
  9.1-3(put,funnel), 9.4(drop), 9.5-10(spray,butter
  etc.), 10(removal), 11(send), 12(push),
  16(concealment), 17(throw), 18.1,2,4(impact),
  22(attach/combine), 23(disassemble),
  43.2,3(roar/flutter), 47.5,7(meander),
  47.8(contiguous location), 48(appear/disappearance
  ), 50(assuming position), 51.1-7(motion),
  53.2(rushing), 59(force), 80(withdraw),
  89(settle), 99(commit)
• Classes that have one of the two patterns
  described earlier 9.1-3, 9.5-10, 10, 11, 17, 48,
  51.2, 51.7, 51.8, 99.

24
Incomplete coverage

• Class-51(verbs of motion) run, dance, skate,
  etc.
• (motion(during(E),Theme)) (Prep(E,Theme,Location))
  
• Mary ran in the forest.
• Mary ran into the forest.
• Mary carried the package in her pocket.
• Class-47 meander etc only stative meaning
• The path meandered through the valley.
• The troops meandered through the valley.

25
What is an argument, what is a sense, what is a
frame?

• Palmer et al. 1999, and Dang et al. 2000
• The bottle floated into the cave.
• The train roared into the station.
• The bottle floated.
• The train roared.
• gt float and roar dont have inherent paths the
  path information has to be adjoined.
• Levin Rappoport Hovav 1995 Roar is polysemous
• VN will only have paths when the constructors of
  VN deemed the path to be inherent.

26
A compositional approach

• Float
• exist(during(E),Theme),Prep(during(E),Theme,Locat
  ion), motion(during(E),Theme)
• Float does not require a path but is compatible
  with one.
• Other verbs require paths, and still others are
  incompatible with them
• This can be done with a simple feature in TAG
  and other frameworks.
• But it is not intuitively clear which verbs the
  authors of VN consider to be in each class, so
  the user has to go through all the (sub)classes
  to find out which bit of information is given or
  not.

27
Outline

• Rationale and aims of the project.
• VerbNet information
• Conclusions

28
Conclusions about VerbNet

• The subcategorization frames that VN handles are
  very incomplete (depends on what was in Levins
  book.)
• VN Semantic structure is a promising piece of
  information to key lexical entailments off but,
  given it is not clear what will be spelled out
  for each class, the user has to go through all
  the classes and subclasses.
• At that point one has done as much work as would
  be required for associating the semantic
  information from scratch with each class or
  subclass.
• Reusability of semantic information doesnt seem
  optimal.

29
What can be done?

• In this particular case
• systematic study of how prepositional phrase
  information and verb information combines in
  change of location expressions.
• Better understanding of what information should
  be contained in subcategorization frames (has
  translating for FrameNet to VerbNet and back and
  then to something else again taught us something?
• More generally given the means available for
  resource development, most likely not very much
  but
• better documentation would be of some help.
• more joint development???
• or is usability what we should aim for and should
  we just forget about reusability?

30
Some of our rules

• !instantiable(VerbSk,t),   
• vn_semantics(VerbSk,location(start(E),Theme,Sou
  rce)),    
• role(Theme, VerbSk, Arg1),    
• role(Source, VerbSk,  Arg2),   
• new_constant(locate,LocSk) !
• gt  
• new_locate(LocSk,  Arg1,  Arg2, VerbSk,  pos,
  before).
• !instantiable(VerbSk,t),   
• vn_semantics(VerbSk,not(location(start(E),Theme,
  Source))),    
• role(Theme, VerbSk, Arg1),    
• role(Source, VerbSk,  Arg2),   
• new_constant(locate,LocSk) !
• gt  
• new_locate(LocSk,  Arg1,  Arg2, VerbSk,  neg,
  before).
• !instantiable(VerbSk,t),   
• vn_semantics(VerbSk,location(end(E),ThemeRole,T
  oLocRole)), 
• role(ThemeRole, VerbSk, Mover),    

31
Some of our rules

• "verb skolem in negative context
• !uninstantiable(VerbSk,t),   
• vn_semantics(VerbSk,location(end(E),ThemeRole,T
  oLocRole)),   
• role(ThemeRole, VerbSk, Mover),    
• role(ToLocRole, VerbSk,  ToLoc),  
•  new_skolem(locate,LocSk) !
• gt  
• new_locate(LocSk, Mover, ToLoc, VerbSk, neg,
  after).
• !uninstantiable(VerbSk,t),   
• vn_semantics(VerbSk,not(location(end(E),ThemeRol
  e,ToLocRole))),   
• role(ThemeRole, VerbSk, Mover),    
• role(ToLocRole, VerbSk,  ToLoc),   
• new_skolem(locate,LocSk) !
• gt  new_locate(LocSk, Mover, ToLoc, VerbSk,
  pos, after). 

32
Some of our rules

• new_locate(LocSk, Mover, Loc, VerbSk, pos,
  ), 
• cached_hypers(locate, Hypers)
• gt 
• subconcept(LocSk, Hypers), 
• instantiable(LocSk, t), 
• role(Location, LocSk, Loc), 
• role(Theme, LocSk, Mover).
• new_locate(LocSk, Mover, Loc, VerbSk, neg,
  ), 
• cached_hypers(locate, Hypers)
• gt subconcept(LocSk, Hypers), 
• uninstantiable(LocSk, t), 
• role(Location, LocSk, Loc), 
• role(Theme, LocSk, Mover). 
• new_locate(LocSk, Mover, Loc, VerbSk, ,
  before)
• gt 
• temporalRel(startsAfterEndingOf, VerbSk,
  LocSk). 

33
Thanks
34
Annotation

• At this point annotation of corpora and lexical
  resources is necessary for deep language
  understanding
• Formal correspondences can be derived from
  aligned corpora
• Meaning correspondences would need an alignment
  between the text and the world
• In some cases this correspondence can be
  approximated with a text-to-text correspondences
  but not in all, e.g. mapping to knowledge
  representation

35
Annotation

• Corpus annotation running text is annotated,
  e.g. coreference
• Annotation of lexical resources, e.g. verb classes

36
Annotation

• Annotation related to meaning is difficult
• Often the distinctions that need to be made are
  not well understood (e.g. animacy, coreference)
• Often the relation to the applications isnt
  clear (e.g. thematic roles)

37
Thematic Roles

• Thematic role labels are used in linguistic
  theory to represent (some of the) inferences that
  the use of specific verbs (or adjectives or
  deverbal nouns) licenses.
• Example John sank the boat (theme)
• The boat (theme) sank
• The label indicates that (some of the inferences)
  are the same in both cases, namely in both
  sentences the boat changes location.

38
VN and inference

• Thematic roles are entailments.
• Based on the thematic role information in VN we
  should be able to derive (some of the)
  entailments of the verb.

39
Thematic roles as entailments

• Thematic roles are meant to mediate between
  syntactic subcategorization and lexical
  semantics. They need to be charactizable in
  semantic terms, otherwise their syntactic use is
  facuous.
• In principle then, thematic roles can be cached
  in as a set of entailments. For instance,
  because the boat is the theme in John sank the
  boat, we know the answer to What sank as well
  as that to What did John sink? and What was
  sunk?

40
Thematic roles as entailmentsGeneral mapping
theory

• Under a general approach to lexical mapping, the
  same thematic role label would be used for the
  same inference across all verbs.
• For instance, in English, verbs such as kiss and
  hit have both a SUBJECT and an OBJECT. In both
  cases the referent of the SUBJECT does the action
  and the referent of the OBJECT undergoes the
  action.
• This means that both verbs have the same thematic
  role for their SUBJECT and for their OBJECT,
  agent and patient respectively.
• (It is, of course, not assumed that the mapping
  is one-to-one for instance, please has an
  experiencer object)

41
Thematic roles as entailmentsNarrow mapping
theory

• Under a narrow conception of thematic role
  labeling, different verb classes have different
  role labels and generalizations are only possible
  within a verb class (cf. FrameNet)We then have
  roles like buyer, seller,

42
What kind of approach is used in Verbnet?

• The documentation for VerbNet is incomplete and
  shattered (Karin Kippers thesis, some articles,
  a couple of files on the web site, )

43
The characterization of thematic roles in Kipper
(2005)

• Theme used for participants in a location or
  undergoing a change of location.
• Agent generally a human or animate subject.
  Used mostly as a volitional agent, but also used
  in VerbNet for internally controlled subjects
  such as forces and machines.
• Destination end point of the motion, or
  direction towards which the motion is directed.
• Source start point of the motion. Usually
  introduced by a source prepositional phrase
  (mostly headed by from or out of.)
• Location underspecified destination, source or
  place in general introduced by a locative or path
  prepositional phrase.

44
Not everything that moves is a theme

• John sent the package to New York.
• Theme package Destination New York.
• John carried the package to New York.
• Theme the package Destination New York.
• and what about John and Mary in the following?
• John followed Mary to New York.
• VN says agent for John and theme for Mary
• John followed Mary with a telescope.
• ?? John followed Mary to the gate with a
  telescope.

45
Not all themes move

• He lives in Hong Kong.
• Theme with verbs of existence (reside, live,
  loom, )
• The tourists admired the paintings.
• Experiencer Theme
• The children liked that the clown had a red nose.
• Experiencer Theme

46
What is a source, what is a path?

• The horse jumped over the fence.
• Theme Location spatial
• Out of the box jumped a little white rabbit.
• Location path Theme
• The convict escaped from the prison.
• Theme Location path
• No destination or source argument for these
  verbs, but
• The books slid from the table.
• Theme Source

47
Patients and Themes

• Hit throw-17.1-1
• Steve hit the ball from the corner to the garden.
• Agent, Themeconcrete()), Source, Destination
• Hit hit-18.1-1
• Paula hit the ball with a stick.
• Agent, Patientconcrete()), Instrument
• Paula hit the ball from the corner to the center
  of the field with a stick.

48
VN semantics (event structure)

• Shove Agent, Theme, Destination
• Amanda shoved the package to the corner.
• motion(during(E),Theme),
• location(end(E), Theme, Destination),
• cause(Agent, E)

49
VN semantics (event structure)

• Live Theme, Location
• exist(during(E), Theme),
• Prep(E,Theme, Location).
• Admire Experiencer, Theme
• emotional_state(E, Emotion, Experiencer),
• in_reaction_to(E, Theme).

50
VN semantics (event structure)

• deport
• The king deported the general to the isle.
• cause(Agent, E),
• location(end(E), Theme, Destination)
• banish
• The king banished the general to the isle.
• cause(Agent, E),
• location(end(E), Theme, Destination)

51
A simple semantic pattern that seems to work

• send class
• Nora sent the book from Paris to London
• motion(during(E), Theme),
• location(start(E), Theme, Source),
• location(end(E), Theme, Destination),
• cause(Agent, E)
• NP(Agent,),verb, NP(Theme,),
  Prep(any,(src,)), NP(Source,),
  Prep(to,),NP(Destination,)

52
Conclusions about VerbNet

• VN thematic roles need to be combined with verb
  class information but at that point the verb
  class and the subcategorization information can
  do the job without the the thematic role.
• Event structure is more useful if one is after
  probable entailments
• Both only encode a subset of the entailments one
  might be interested in.