A Computational View of Verb Predicates and Semantic Roles

1
A Computational View of Verb Predicates and
Semantic Roles

• Fernando Gomez
• School of Computer Science
• University of Central Florida
• Orlando, Fl 32816
• gomez_at_cs.ucf.edu
• www.cs.ucf.edu/gomez

2
Description of the Problem

• Design and implement a program that takes as
  input any English sentence and, for every clause
  in the sentence,
• Determines the verb meaning, the semantic roles,
  adjuncts, attach PPs, solves the senses of many
  nouns in the sentence, and also resolves deverbal
  nominalizations.

3
The Role of WordNet

• WordNet provides two major resources for
  defining the predicates
• a) A General Ontology for English Nouns
• Used in the selectional restrictions of the
  predicates
• b) A Classification of English Verb into Classes
• Used in identifying generic predicates whose
  definitions would apply to many verbs under that
  class

4
An example of WordNet Verb Class travel

• Sense 1
• travel, go, move, locomote -- (change location
  move, travel, or proceed)
• gt go around, spread, circulate -- (of
  information)
• gt carry -- (cover a certain distance
  or advance beyond, as of a ball in golf "The
  drive carried to the green")
• gt ease -- (move gently or carefully "He
  eased himself into the chair")
• gt whish -- (move with a whish)
• gt seek -- (go to or towards "a liquid
  seeks its own level")
• gt whine -- (move with a whining sound
  "The bullets were whining past us")
• gt fly -- (be dissipated "Rumors and
  accusations are flying")
• gt ride -- (move like a floating object
  "The moon rode high in the night sky")
• gt come -- (cover a certain distance "She
  came a long way")
• gt ghost -- (move like a ghost "The
  masked men ghosted across the moonlit yard")
• gt travel -- (undergo transportation, as
  in a vehicle)
• gt fly -- (travel in an airplane "she
  is flying to Cincinnati tonight" "Are we driving
  or flying?")
• gt hop -- (informal travel by means
  of an aircraft, bus, etc. "She hopped a train to
  Chicago" "He hopped rides all over the country")
• gt ride -- (be carried or travel on or
  in a vehicle "I ride to work in a bus" "He
  rides the subway downtown every day")
• gt chariot -- (ride in a chariot)
• gt bicycle, cycle, bike, pedal,
  wheel -- (ride a bicycle)

5
The Syntax of Roles

• (role ( (ltslrgt) (ltsrsgt) ... (ltslrgt) (ltsrsgt)))
• where
• ltrolegt stands for any Semantic Role
• ltslrgt stands for one or more Selectional
  Restrictions
• and
• ltsrsgt stands for one or more Grammatical
  Relations/Syntactic Relations
• Example (to-loc(location) (obj)
• (physical-thing)
  ((prep to)))

6
Examples of Grammatical Relations

• Subj Subject of verb
• Obj First Postverbal NP
• Obj2 Second Postverbal NP
• Subj-if-obj Subject of a verb that has also an
  obj
• Prep PP

7
Examples of Grammatical Relations

• CP Any Complement Phrase
• CP-S VP_S clause
• CP-INF VP_inf
• CP_ING VP_ing
• PREP-CP CP clause introduced by a preposition

8
A Generic Predicate Communicate

• COMMUNICATE
• (IS-A (INTERACT))
• (WN-MAP (COMMUNICATE2))
• (AGENT (HUMAN-AGENT ANIMAL ) (SUBJ))
• (THEME (COMMUNICATION (POSSESSION PHYSICAL-THING
  STATE-R)
• 
  THING) (CP OBJ OBJ2)
  
  
  
  
  
  
• (COMMUNICATION THING) ((PREP
  ABOUT OF))
• (COMMUNICATION ABSTRACTION)
  ((PREP ON)))
• (RECIPIENT (HUMAN-AGENT ANIMAL) (OBJ (PREP
  WITH TO)))
• (FORM-OR-MEDIUM-OF-EXPRESSION(WRITTEN-COMMUNICATI
  ON SPEECH-ACT CREATION) (SUBJ
  (PREP IN)))

9
"She briefed the president."

• (Clause CL11
• (SUBJ ((PRON SHE))
• (PERSON SHE) lt(AGENT)gt
• )
• (VERB BRIEF ltTRANS-INFOR(BRIEF1)
  supported by 2 SRsgt )
• (OBJ ((DFART THE) (NOUN PRESIDENT))
• (LEADER PRESIDENT1 PRESIDENT2
  PRESIDENT3 PRESIDENT4
• PRESIDENT5) lt(RECIPIENT)gt
• ))
• TRANS-INFOR
• COMMUNICATE
• INTERACT
• ACTION
• ROOT

10
Criteria for creating subpredicates

• The differentia between a predicate and its
  subpredicates are given by one or more of the
  following
• Different selectional restrictions for the
  semantic roles
• Different syntactic relations for the semantic
  roles
• Specific sets of inferences associated with the
  subpredicates

11
Definitions for Some Small Classes of Communicate2

• ADVISE
• (IS-A(COMMUNICATE))
• (WN-MAP(ADVISE1))
• (THEME(THING) (CP (PREP ABOUT))
• ((PHYSICAL-THING) THING) ((PREP
  ON)))
• (RECIPIENT(HUMAN-AGENT ANIMAL) (OBJ))
• SMILE
• (IS-A(GRIMACE))
• (WN-MAP(SMILE1))
• (RECIPIENT(HUMAN-AGENT ANIMAL) ((PREP AT TO)))
• (THEME(NIL)(NIL))

12
Something We Missed

• The senator advised against the war.
• The doctor advised complete rest.
• Please advise me of the cost.

13
A Small Verb Subclass

• Sense 1
• smile -- (change one's facial expression by
  spreading the lips, often to signal pleasure)
• gt dimple -- (produce dimples while
  smiling "The child dimpled up to the adults")
• gt grin -- (to draw back the lips and
  reveal the teeth, in a smile, grimace, or snarl)
• gt beam -- (smile radiantly express joy
  through one's facial expression)
• gt smirk, simper -- (smile affectedly or
  derisively)
• gt sneer -- (smile contemptuously)

14
A Subpredicate of Transfer of Information

• MISINFORM (MISINFORM1)
• HIDE-INFORMATION
• CONCEAL-INFORMATION
  (CONCEAL2)
• COVER-HIDE-INFORMATION
• COVER-UP-MISINFORM
• SWEEP-UNDER-THE-RUG
• LIE-TO-SOMEBODY (LIE5)
• EXAGGERATE (OVERSTATE1)
• BLOW-EXAGGERATE
• TWIST-CHANGE-THE-MEANING (TWIST8)
• CONTORT-DISTORT (CONTORT?)

15
A Subpredicate of Transfer of Information

• TEACH (TEACH1)
• HAMMER-IN (HAMMER_IN1)
• HAMMER-TEACH
• LECTURE (LECTURE1)
• EDUCATE (EDUCATE1 TRAIN2 EDUCATE3)
• BRING-UP (REAR2)
• TRAIN-SOMEBODY (TRAIN1
  TRAIN2)
• CULTIVATE-KNOWLEDGE-WISDOM
  (CULTIVATE3)
• EDUCATE-PEOPLE
• PREPARE-SOMEBODY-FOR-SCHOOL
• INDOCTRINATE (INDOCTRINATE1)
• INFECT-INDOCTRINATE
• POISON-INDOCTRINATE
• TRAIN-ANIMAL (TRAIN1)
• IMPRESS-ON-SOMEONE
• INSTILL-PUT-IDEAS
• INSTRUCT-TEACH
• SHOW-TEACH

16
A Subpredicate of Transfer of Information

• CRITICIZE (CRITICIZE1)
• DISPARAGE (DISPARAGE1)
• ATTACK-VERBALLY (ATTACK2)
• CURSE-BLASPHEME (CURSE1
  CURSE2)
• SWEAR-BLASPHEM
• VILIFY-VITUPERATE
  (VILIFY1)
• REBUKE (REBUKE1)
• CHASTIZE (CHASTIZE1)
• DENOUNCE (DENOUNCE1 DENOUNCE3
  DENOUNCE4)
• CONDEMN-DENOUNCE
• PICK-ON-CRITICIZE
• TEAR-APART-CRITICIZE

17
Semantic Roles and Predicate Classes

• Semantic Roles Depend on the Generic Predicate
  for Each Predicate Class
• The Generic Predicate Determines the Meaning and
  Number of the Semantic Roles (Gomez, 1998)

18
Examples of Generic Predicate Classes

• Change-Location (Agent, To-Loc, From-Loc,
  Distance, Instrument, At-Speed)
• Cause-to-Change-Location (Agent, Theme, Source,
  Goal,
• 
  Inanimate-Cause, Instrument)
• Transfer-of-Possession (Agent, Theme, From-Poss,
  To-Poss)
• Cause-Change-of-State (Agent, Theme,
  Beginning-State, Ending-State,
• 
  Inanimate-cause)
• Prevent (Agent, Theme, Event-Prevented,
  Recipient, Inanimate-Cause)
• Judge (Agent, Theme, Recipient)
• Permit (Agent, Theme, Recipient, Inanimate-Cause)

19
Hierarchy of Semantic Roles

• from-person gt beginning-state
  (BREAK-OFF-FROM-SOMEBODY)
• in-court-tribunal gt at-loc (CHALLENGE-SOMETHIN
  G)
• event-prevented gt theme (PREVENT)
• at-inquiry-investigation gt at-activity
  (CLEAR-SOMEBODY-OF-BLAME-GUILT)
• cure-agent gt inanimate-cause (CURE-A-DISEASE)
• work-place gt at-loc (DO-SERVICE)
• from-organization-or-activity gt
  beginning-state (RETIRE-FROM-AN-ACTIVITY)

20
Semantic Interpretation Algorithm

• The interpretation algorithm is activated after a
  sentence is parsed.
• The parser does not resolve structural ambiguity.
• The determination of verb meaning and semantic
  roles is interdependent.
• A predicate explains a syntactic relation if it
  has a semantic role realized by that syntactic
  relation.
• The predicate that has the most semantic roles
  realized is selected as the meaning of the verb.

21
Illustration of the Algorithm

• P1 leave-a-place
• P2 leave-an-organization
• P3 abandon-somebody
• P4 leave-a-vehicle
  
• leave P5 leave-behind
• P6 leave-give
• .
• .
• .
• Pi
• She left for Texas on a plane. She left a fortune
  to her daughter.
• He left Texas. She left school. He left with his
  friends.

22
Problems with the WordNet Verb Classes as Relate
to Predicates

• Verb forms within a class may realize their
  semantic roles by different
• a) Syntactic Relations
• and/or
• b) Selectional Restrictions

23
Problems with the selectional restrictions

• Sense 1
• reach, attain, make, hit, arrive at, gain --
  (reach a destination, either
• real or abstract "We hit Detroit by noon" "The
  water reached the doorstep"
• gt catch up -- (reach a the point where
  one should be after a delay "I caught up on my
  homework")
• gt come back -- (even the score, in
  sports)
• gt scale, surmount -- (reach the highest
  point of "We scaled the Mont Blanc")
• gt breast -- (reach the summit "They
  breasted the mountain")
• gt access, get at -- (reach or gain access
  to)
• gt peak, reach a peak -- (to reach the
  highest point attain maximum intensity,
  activity "That wild, speculative spirit peaked
  in 1929.")
• gt crest -- (reach a high point "The
  river crested last night")
• gt bottom out -- (reach the low point)
• gt make -- (reach in time "We barely made
  the plane")

24
Problem with the Syntactic Relations

• Sense 1
• accuse, impeach, incriminate, criminate -- (bring
  an accusation against level a charge against
  "He charged the man with spousal abuse")
• gt reproach, upbraid -- (utter a reproach
  to "The president reproached the general for his
  irresponsible behavior")
• gt arraign -- (accuse of a wrong or an
  inadequacy)
• gt impeach -- (charge with a crime or
  misdemeanor)
• gt recriminate -- (return an accusation
  against someone or engage in mutual accusations
  charge in return)
• gt charge, lodge, file -- (file a formal
  charge against "The suspect was charged with
  murdering his wife")
• gt impeach -- (charge with an offense
  or misdemeanor "The public officials were
  impeached")

25
Defining Predicates for Individual Verbs with
High Polysemy

• The definition is identical as for defining
  predicates for verb classes.
• However,
• The order in which the predicates are defined is
  relevant because the algorithm prefers them in
  the order in which they are defined.

26
An Example of Predicates for an Individual Verb

• ADDRESS
• (ADDRESS-AN-ENVELOPE
• (IS-A(LABEL-SOMETHING))
• (WN-MAP(ADDRESS3))
• (THEME(ENVELOPE1)(OBJ))
• (THEMATIC-RULE(REQUIRE(THEME )))
• )
• (ADDRESS-SOMEONE
• (IS-A(SPEAK-TO-SOMEBODY))
• (WN-MAP(ADDRESS1)(ADDRESS2))
• (AGENT(HUMAN-AGENT ANIMAL)(SUBJ))
• (RECIPIENT(HUMAN-AGENT ANIMAL)(OBJ))
• )

27
A new address

• (ADDRESS-A-PROBLEM-TASK-SITUATION
• (WN-MAP(ADDRESS6))?
• (IS-A(DEAL-WITH-A-PROBLEM-TASK-SITUATION))
• (THEME(DIFFICULTY2 DIFFICULT3 CHALLENGE1
  ACTION )(OBJ))
• )
• SHE ADDRESSED MANY IMPORTANT TOPICS IN HER
  BOOK.
• (ADDRESS-DISCUSS
• (IS-A(DISCUSS-ABOUT))
• (WN-MAP(ADDRESS7))
• (THEME((HUMAN-AGENT) THING) (OBJ))
• )

28
An Example of Predicates for an Individual Verb

• (ADDRESS-A-SPEECH-WRITTEN-COMMUNICATION-TO-SOME
  BODY
• (IS-A(TRANS-INFOR))
• (WN-MAP(ADDRESS5)) ?
• (AGENT(HUMAN-AGENT ANIMAL)(SUBJ-IF-OBJ))
• (THEME(COMMUNICATION)(OBJ))
• (THEMATIC-RULE(REQUIRE(THEME )))
• )
• ALL ICELANDERS ARE ADDRESSED BY THEIR FIRST
  NAMES.
• (ADDRESS-GREET-SOMEONE
• (IS-A(NAME-SOMETHING))
• (WN-MAP(ADDRESS4))
• (RECIPIENT(HUMAN)(OBJ))
• (NAME-OF(LANGUAGE_UNIT1)((PREP BY AS))
• (PERSON) ((PREP AS))) )

29
Grounding the Ontology on the Semantic
Interpretation Algorithm (Gomez, 2001, 2003)

• The role of the ontology is essential because if
  the ontological categories are wrong, the
  selectional restrictions in the predicates will
  be also wrong.
• We did not proceed by looking into the
  upper-level ontology to find out which categories
  may require changes. But,
• The testing of the predicates determined for us
  which ontological categories may require changes.

30
Illustrations of Ontological Changes

• Define Selectional Restriction for hide 2 in WN.
• The fish hides in a crevice.
• Define Selectional Restrictions for burn 1 in WN.
• She burned the letter.
• Define Selectional Restrictions for flow 2 in WN.
• Blood flowed from the wound.

31
Status of the Work

• We have mapped 95 of WN verb classes into
  predicates.
• We have defined over 3000 predicates.

32
Panorama of the Upper-level Ontology of Predicates

• Cause-change-of-state (609 subpredicates)
• cause-change-of-state-of-animal-being (140)
• arouse-feelings-emotions (52)
• cause-to-act (19)
• injure-hurt-somebody-or-oneself (18)
• increase (31)
• improve (19)
• worsen (10)
• terminate (16)
• physical-change-of-state (14)
• solidify
• liquefy
• cause-change-of-integrity (22),
• transform (9) and others.

33
Cause-to-Change-Location

• Cause-to-change-location (379)
• put (75)
• remove (53)
• transport (23)
• propel (20)
• connect (22)
• flow (12)
• pull (9)
• push (9)
• send (9)

34
Change-Location and Interact

• Change-Location (238)
• walk (14)
• hike
• march
• sneak-walk
• enter (10)
• leave-a-place (11)
• arrive-to-a-place (10)
• Interact (372)
• communicate (243)
• treat-an-animal-or-human (26)
• treat-unjustly-somebody (9)
• behave (9)
• join-a-group-or-a-human (9)

35
Transfer of Something and Make-Or-Create-Something

• Transfer-of-something (293)
• transfer-of-possession (231)
• give (31)
• get (132)
• Make-or-create-something (144)
• initiate-something (30)
• create-art (28)
• produce (10)
• prepare-something (10)

36
Judge

• Judge (182)
• pass-a-negative-judgment-on (36)
• disapprove
• oppose-ideas
• oppose-people
• put-value-to-something (12)
• praise (7)
• accept-admit-a-fact (30)
• confirm-corroborate (11)
• deny-something-to-somebody (14)
• accuse (11)

37
Experience, Think and Decide

• Experience (110)
• feel-a-state-or-emotions (25)
• perceive (21)
• like-something (36)
• experience-event-state-abstraction (11)
• Think (115)
• analyze (31)
• plan (13)
• reason-conclude (9)
• associate (7)
• Decide (92)
• exert-control-over (65)
• restrain (22)
• manage (23)

38
Touch and Spend-Something

• Touch (55)
• handle-operate (18)
• hit-something (24)
• Spend-Something (52)
• ingest (31)
• Eat

39
Other Unique Predicate Classes

• Prevent (45)
• Move-body-position (33)
• Know (31)
• Fail-to-do-something (33)
• Fight (31)
• Expel-substance-from-the-body (23)
• Do-act (27)
• Appoint-somebody (15)
• Allow-something (21)
• Support-something (24)
• Succeed (23)
• Utilize (11)
• Stative-Predicates (223)
• be-at-a-place
• include

40
Testing of the Predicates

• We have tested 400 verbs and produced a small
  corpus of 500 interpreted sentences from The
  World Book Encyclopedia, (WorldBook, Inc.
  Chicago). (This corpus is available in my
  homepage)
• For verbs having 10 senses or more, the algorithm
  selected correctly the meaning of the verb in 85
  of the cases.
• For verbs having less than 10 senses, the
  algorithm selected the correct sense of the verb
  in 92 of the cases.
• If the predicate is selected correctly, the
  semantic roles are correctly determined 97 of
  the cases.

41
Conclusions

• We have presented methods in lexical semantic to
  define predicates for English verbs.
• The method uses WordNet noun ontology for the
  selectional restrictions in the semantic roles of
  the predicates, and
• It also uses WordNet verb classes to define
  generic predicates that apply to a large class of
  verbs.
• We have provided definitions for over 3000
  predicates, and mapped 95 of WordNet verb
  classes into predicates.

42
Conclusions (continuation)

• An algorithm that uses the predicate definitions
  has been designed and implemented.
• The algorithm is used to test and refine the
  definition of the predicates.
• The algorithm has provided essential information
  to reorganize the upper-level ontology of
  WordNet.
• By using the algorithm and the predicates, we
  have given some steps to automatically produce
  semantic tagged corpora.

43

• Some References
• C. Fellbaum (1998) "A Semantic Net of English
  Verbs" In WordNet An electronic Lexical Database
  and some of its applications, Fellbaum, C.
  (editor) MIT Press, 1998.
• F. Gomez (1998) "Linking WordNet Verb Classes to
  Semantic Interpretation, In the COLING-ACL
  Workshop on Usage of WordNet in NLP, U. of
  Montreal.
• F. Gomez (2001) "An Algorithm for Aspects of
  Semantic Interpretation Using an Enhanced
  WordNet," In 2nd Meeting of the North American
  Chapter of the Association for Computational
  Linguistics, NAACL-2001, CMU.
• F. Gomez (2001) "Grounding the Ontology on the
  Semantic Interpretation Algorithm", CS-TR-01-01,
  Feb-2001. Also to appear in the 2nd International
  Conference in WordNet, Jan-04
• G. Miller (1998) "Nouns in WordNet," in WordNet
  An electronic Lexical Database and some of its
  applications", Fellbaum, C. (editor) MIT Press,
  1998.
• Levin, B. English Verb Classes and Alternations
  A Preliminary Investigation University of Chicago
  Press, 1993, Chicago.
• Pinker, S. Learnability and Cognition, MIT Press,
  1989, Cambridge, Mass.
• Pritchett, B. L. Grammatical Competence and
  Parsing Performance, The University of Chicago
  Press", 1992. Chicago,Illinois.
• Grimshaw, J. Argument Structure, MIT Press, 1990,
  Cambridge, Mass.