FrameNet, PropBank, VerbNet
Rich Pell
FrameNet, PropBank, VerbNet




When syntactic information is not enough
Lexical databases
Annotate a natural language corpus with semantic
information
Largely manual classification efforts
Outline

FrameNet




PropBank



Overview and Demo
Applications
VerbNet



Frame Semantics
Overview and Demo
Applications
Levin Classes
Frame Demo
Conclusion
FrameNet
Frame Semantics



Sell (v) – to exchange an item for money or its equivalent
Vce (n) – estimator of reactivity delta due to voids in
moderator
Definition of a word is useless without knowledge
relating to that word:


Entities involved - buyer, seller, item, money
Relationships between those entities:




Buyer gives money to seller
Seller gives item to buyer
Buyer believes value of item >= monetary amount
Seller believes value of item <= monetary amount
Semantic Frame



Description of an event, relation, or entity and its
participants
Captures the ‘essential knowledge’ of a given word sense
Developed by Charles Fillmore
FrameNet Overview






Attempt to represent frame semantics in a human and
machine-readable database
Developed by Charles Fillmore at Berkeley’s International
Computer Science Institute
Founded in 1997
Funded by National Science Foundation and DARPA
Freely available via web interface or download
https://framenet.icsi.berkeley.edu/fndrupal/
FrameNet Overview






Set of semantic frames
Composed of frame elements (FEs) – roles within the
frame
Words that evoke this frame are called lexical units(LUs) –
represent a sense of a given word
Frame: Commerce_sell
FEs: buyer, seller, item, money, place, reason…
LUs: auction.v, retail.v, vend.v…
Frames


Definition
Core/non-core frame elements



Definition and examples
Frame-frame relations
Lexical Units
Frame-Frame Relations

Inheritance – IS-A relation




Using – child frame presupposes parent frame as background



Speed presupposes Motion
No one-to-one correspondence between FEs
Subframe – child frame is subevent of complex event
represented by parent


Child frame is subtype of parent frame
Each frame element in parent has corresponding frame element in
child
Revenge inherits from Rewards_and_punishments
Criminal_process -> Arrest, Arraignment, Trial, Sentencing
Perspective-on – one frame provides some perspective on
(perspectivizes) another frame

Commerce_goods_transfer provides perspective on Commerce_sell
Text Annotation





[Seller Bob] auctioned [Goods the clock] [Buyer to John]
[Item Colgate’s stock] rose [Difference $3.64][Final_value
to $49.94]
reduction [Item of debt levels][Value_2 to $665
million][Value_1 from $2.6 billion]
[Sleeper They][Copula were]asleep[Duration for hours]
He took a packet of Woodbines out of the breast pocket
of [Wearer his][Material cotton][Garment shirt] and lit
one.
Development



Characterize frames
Find words that fit the frames (lexical units)
Extract sample sentences





British National Corpus (editorials, sermons, textbooks,
advertisements, novels, sermons)
Linguistic Data Consortium (US newswire texts)
American National Corpus
~200 million words
Annotate selected examples
Progress

1000 linked semantic frames comprising:



10,000 lexical units
170,000 manually annotated sentences
Ports to other languages

Spanish, German, Chinese, Japanese
Uses







Semantic role assignment
Natural language understanding
Machine translation
Part of speech tagging
Textual entailment
Information extraction
NLP applications where a syntactic parse will not suffice
PropBank
PropBank



Adds a semantic layer to Penn Treebank
Attempts to capture accurate predicate-argument
structure by annotating predicates and the semantic roles
of their arguments
Annotates predicates (verbs) and their arguments:






John broke the window -> broke(arg0 = John, arg1 = the
window)
The window broke -> broke(arg1 = the window)
Developed in 2001 at the University of Pennsylvania
Martha Palmer, Paul Kingsbury
Free, open-source, downloadable
http://verbs.colorado.edu/~mpalmer/projects/ace.html
PropBank Structure



PropBank is a set of frame files
Each frame file contains one or more PropBank verb
senses (aka frameset or roleset ID)
Each verb sense is annotated with:



Semantic roles for each argument of a predicate
Examples
Links to other lexical tools (FrameNet,VerbNet)
PropBank Arguments

Standardized as much as possible







Arg0 = agent
Arg1 = patient
Arg2 = instrument/attribute
Arg3 = starting point/attribute
Arg4 = ending point
ArgM = modifier
Obama met him privately in the White House, on Thursday.






Rel: met
Arg0: Obama
Arg1: him
ArgM-MNR: privately
ArgM-LOC: in the White House
ArgM-TMP: on Thursday
PropBank Example
…the campaign is drawing fire from anti-smoking
advocates…
Arg0:
the campaign
Rel:
drawing
Arg1:
fire
Arg2-from: anti-smoking advocates
PropBank Example
<predicate lemma="sell">
<roleset id="sell.01" name="commerce: seller" vncls="13.1-1">
<roles>
<role descr="Seller" n="0">
<vnrole vncls="13.1-1" vntheta="Agent"/></role>
<role descr="Thing Sold" n="1">
<vnrole vncls="13.1-1" vntheta="Theme"/></role>
<role descr="Buyer" n="2">
<vnrole vncls="13.1-1" vntheta="Recipient"/></role>
<role descr="Price Paid" n="3"/>
<role descr="Benefactive" n="4"/>
</roles>
<example name="intransitive">
<text>
They-1 have *trace*-1 to sell when things look like they're falling.
</text>
<arg n="0">*trace*</arg>
<rel>sell</rel>
<arg f="TMP" n="M">when things look like they're falling</arg>
</example>
<example name="Ergative">
<text>
A painting by August Strindberg sold at auction in Stockholm.
</text>
<arg n="1">A painting by August Strindberg</arg>
<rel>sold</rel>
<arg f="LOC" n="M">at auction</arg>
<arg f="LOC" n="M">in Stockholm</arg>
</example>
Differences From FrameNet




Verb-specific
Each verb is its own predicate
Closer to syntactic parse
More thorough but simpler annotation of corpus
PropBank Progress





3500 verbs annotated
Work on translating to Dutch, Arabic
Semantic role labeling
Knowledge discovery
Semantic parsing
VerbNet
VerbNet



Lexicon of English verbs
Groups verbs based upon shared syntactic behavior
5800 verbs in 270 verb classes




Based on Levin classes and their extensions
Developed by Karin Kipper-Schuler at University of
Pennsylvania via NSF and DARPA grants
Free, open source, downloadable
http://verbs.colorado.edu/~mpalmer/projects/verbnet.htm
l
Levin Classes




English Verb Classes and their Annotations, Beth Levin, 1993
Syntactic behavior of a verb is based upon its meaning
Possible to syntactically group verbs into classes based
upon how they interact with specific
objects/prepositions/subjects and expect them to have
some semantic similarity
e.g. Locative alternation – involves moving something into
or onto a location


Verbs of placement and covering
Scatter, pump, hang, drizzle, cram, load
VerbNet Roles


Groups verbs based upon Levin classes
Add semantic role labels to Levin classes, e.g.







Agent – actor in an event who carries out the event
Theme – undergoer that is central to event or state that does
not have control over the way the event occurs
Destination – goal that is a concrete, physical location
…
23 total
Illustrate the “who what how when where” information
contained in a sentence
Analogous to FrameNet’s frame elements or PropBank’s
numbered arguments
VerbNet Classes



Set of member verbs
Thematic roles used in predicate-argument structure of
verbs in the class
Selectional restrictions on the roles



“Sam drank a coffee.”
“Sam drank a car.”
Set of frames:




Brief description
Example
Syntactic description
Set of semantic predicates, includes temporal function
indicating when a predicate is true
VerbNet applications









Verbs typically convey the main idea of a sentence
Maps the syntactic nature of PropBank
predicate/argument parses into a richer semantic context
Machine translation
Document classification
Word sense disambiguation
Semantic role labeling
3D animation (parameterized action representations)
Planning
Automatic verb acquisition
Automatically Extending VerbNet



Semantic information for several verbs at a time captured
in VerbNet classes
Can automatically add new candidate verbs to a class by
testing against pre-defined class specifications
Removes need for exhaustive manual encodings
Automatically Extending VerbNet

Apply k-means clustering to some other resource:






PropBank
WordNet
FrameNet
Observe the clusters to see if they correspond to any
VerbNet class
If so, do they contain any verbs not in the existing
VerbNet class?
Able to add 47 verbs
Summary

FrameNet, PropBank,VerbNet all annotate an NL corpus
with semantic information:





FrameNet – defines a set of semantic frames annotating
additional semantic information needed to capture meaning of
a word
PropBank – annotates propositions and their arguments in a
structured fashion
VerbNet – groups verbs into syntactically and semantically
similar classes
All are used when a syntactic parse is not enough
Highly linked:


Unified Verb Index - http://verbs.colorado.edu/verb-index/
SemLink - http://verbs.colorado.edu/semlink/
Questions
References
VerbNet Guidelines, http://verbs.colorado.edu/verb-index/VerbNet_Guidelines.pdf
Palmer, M. 2009. Semlink: Linking PropBank,VerbNet and FrameNet. Proceedings of the Generative Lexicon
Conference. Sept. 2009, Pisa, Italy: GenLex-09.
2012 Data Format Specifications for English PropBank, http://verbs.colorado.edu/~mpalmer/projects/ace/EPB-dataformat.txt.
M. Palmer et al, “English PropBank Annotation Guidelines,” 2012, http://verbs.colorado.edu/~mpalmer/projects/ace/EPBannotation-guidelines.pdf
Karin Kipper, Anna Korhonen, Neville Ryant, and Martha Palmer. Extending VerbNet with Novel Verb Classes. Fifth
International Conference on Language Resources and Evaluation (LREC 2006). Genoa, Italy. June, 2006.
Karin Kipper, Anna Korhonen, Neville Ryant, and Martha Palmer. Extensive Classifications of English verbs. Proceedings
of the 12th EURALEX International Congress. Turin, Italy. September, 2006.
Paul Kingsbury and Karin Kipper. Deriving Verb-Meaning Clusters from Syntactic Structure..Workshop on Text
Meaning, held in conjunction with HLT/NAACL 2003. Edmonton, Canada, May 2003.
Karin Kipper-Schuler,VerbNet: a Broad-Coverage, Comprehensive Verb Lexicon,” Dissertation, University of
Pennsylvania, 2005.
Michael Ellsworth et al, “FrameNet II: Extended Theory and Practice,” 2010,
https://framenet2.icsi.berkeley.edu/docs/r1.5/book.pdf.