A Hierarchical Unifica.on of LIRICS and VerbNet Seman.c Roles
Claire Bonial, William Corvey, Martha Palmer, Volha V. Petukhova, and
Harry Bunt 2011
Overview and Purpose
• Compares thema.c roles of VerbNet (VN) and thema.c roles of the Linguis.c InfRastructure for Interoperable ResourCes and Systems (LIRICS)
• Replace exis.ng VN seman.c roles and develop a standard set of thema.c roles that would be appropriate for a variety of NLP applica.ons
Proposal: A unified set of seman.c roles that will be
incorporated into VN
• Map more easily to an ISO standard set • Arranged hierarchically based on seman.c feature inheritance and feature constraints
LIRICS • Seman.c roles should be defined – As seman.c categories (not syntac.c or lexical structures)
– By dis.nc.ve seman.c proper.es – As not restricted to only a few specific verb (/noun/adjec.ve) classes
– As rela.onal no.ons that link par.cipants to an event • Inten.onal ac.on? Affected by other par.cipants? Exist through the event?
LIRICS
• Set of 29 roles – 11 central roles (agent, theme, pa.ent) – 10 adjunct roles (.me, loca.on, manner) – 8 subroles for .me and loca.on (dura.on, frequency, path)
LIRICS Example: theme role
• Essen.al to the event taking place, does not have control over the way the event occurs • Not structurally changed by the event
• In a fixed posi.on/condi.on throughout the state
• Essen.al to the state being in effect • Not as essen.al as the pivot role
LIRICS • Granularity – Low-‐level roles inherit all the proper.es of high-‐level roles and have an addi.onal feature
– Result: a shallow hierarchy for several seman.c roles
– Example: • Time: Ini)al_Time, Final_Time, Frequency, and Dura)on
VerbNet
• Describes the sets of diathesis alterna.ons that are compa.ble with each verb in the lexicon
• Example:
Break Appear
Inchoa.ve varia.on
The window broke
A rabbit appeared out of the magician’s hat
Causa.ve varia.on
The boy broke the window
*The magician appeared a rabbit out of his hat
VerbNet
• Fundamental assumption: syntactic frames that are compatible with a particular verb are a reflection of the underlying semantics
• Thematic roles are assigned to each syntactic argument in a given verb class
• Use of semantic predications denoting relationships between participants and events
VerbNet
• Example: I put the book on the table Syntac.c Representa.on Seman.c Representa.on
NP V NP PP MOTION(DURING(E), THEME) Agent V Theme Des.na.on NOT(PREP: on (START(E), THEME, DESTINATION)) PREP: on (END(E), THEME, DESTINATION) CAUSE(AGENT, E)
Current State of VN makes use of:
• Commonly used, coarse-grained roles like those of LIRICS (e.g. agent)
• Roles that are specific to certain classes of events, which are intended to convey key semantic components of some verb classes (e.g. topic)
• Roles that are partially syntactically motivated (e.g. predicate)
• Roles that are distinguished by internal properties of the participant (e.g. +/- animate)
LIRICS only makes use of roles that are:
• Not restricted to specific verb classes • Not linked to particular syntactic structures • Not related to internal properties of
participants
Old VN role New VN role LIRICS role Actor Agent Agent Actor 1 Agent Agent Actor 2 Co-‐Agent Partner Agent Agent Agent Asset Asset Amount
A_ribute A_ribute A_ribute Beneficiary Beneficiary Beneficiary
Cause Cause Cause Des.na.on Des.na.on Final_Loca.on Experiencer Experiencer Pivot
Extent Extent Amount/Distance Instrument Instrument Instrument Loca.on Loca.on Loca.on Material Material Source Pa.ent Pa.ent Pa.ent Pa.ent 1 Pa.ent Pivot Pa.ent 2 Co-‐Pa.ent Pa.ent Predicate -‐ -‐ Product Result/Product Result
Proposi.on -‐ -‐ Recipient Recipient Goal Source Source/Ini.al_Loca.on Source/
Ini.al_Loca.on S.mulus S.mulus Theme Theme Theme Theme Theme 1 Theme Theme Theme 2 Co-‐Theme Theme Time Time Time Topic Topic Theme
Predicate -‐ -‐ Product Result/Product Result
Proposi.on -‐ -‐ Recipient Recipient Goal Source Source/Ini.al_Loca.on Source/
Ini.al_Loca.on
S.mulus S.mulus Theme
Theme Theme Theme
Theme 1 Theme Theme
Theme 2 Co-‐Theme Theme
Time Time Time
Topic Topic Theme
Roles specific to VN
• Example: • Topic – Fine-‐grained – Specific to verbs of communica.on – More likely to be realized in the form of a complement clause
• Theme – Coarse-‐grained – More likely to be realized in the form of a noun phrase
Hierarchy
• Coarse-‐grained roles: for tasks that require a roleset that has the broadest coverage across all verbs
• Fine-‐grained roles/Class specific roles: for tasks that benefit from info that helps to dis.nguish classes of verbs
Revised Roleset
• Roles: Par.cipant, Actor, Undergoer, Pivot • Roles adapted or revised from LIRICS: Cause, Agent, Instrument, Theme, Pa.ent, A_ribute, Beneficiary, Place, Source, Ini.al_Loca.on, Goal, Time, Ini.al_Time, Final_Time, Frequency, Dura.on
• Roles specific to certain event types: Co-‐Agent, S.mulus, Co-‐Theme, Topic, Co-‐Pa.ent, Experiencer, Loca.on, Material, Des.na.on, Recipient, Result, Product, Value, Extent, Asset
Examples • He talked about poli.cs LIRICS: HeAGENT talkedRELATION about_poli.csTHEME
VN 1 (coarse-‐grained): HeAGENT talkedRELATION about_poli.csTHEME
VN 2 (fine-‐grained): HeAGENT talkedRELATION about_poli.csTOPIC • He sent the le_er to Mary LIRICS: HeAGENT sentRELATION the_le_erTHEME to_MaryGOAL VN 1 (coarse-‐grained): HeAGENT sentRELATION the_le_erTHEME to_MaryGOAL
VN 2 (fine-‐grained): HeAGENT sentRELATION the_le_erTHEME to_MaryRECIPIENT
Examples • John collaborated with Paul on the task. LIRICS: JohnAGENT collaboratedRELATION with_PaulPARTNER on_the_taskTHEME
VN: JohnAGENT collaboratedRELATION with_PaulCO-‐AGENT on_the_taskTHEME
• The tourists admired the pain.ngs. LIRICS: The_touristsPIVOT admiredRELATION the_pain.ngsTHEME VN 1 (coarse-‐grained): The_touristsPATIENT admiredRELATION the_pain.ngsCAUSE
VN 2 (fine-‐grained): The_touristsEXPERIENCER admiredRELATION the_pain.ngsSTIMULUS