Transcript Verbnet/Framenet

Steven Schoonover






What is VerbNet?
Levin Classification
In-depth look at VerbNet
Evolution of VerbNet
What is FrameNet?
Applications





Lexical resource that organizes English verbs
into different classes
Describes verbs based on their syntactic and
semantic properties
Originally put together in 2000 by Martha
Palmer and Karin Kipper (released in 2005)
Based on the Levin Classification of 1993
Maps to things including WordNet, Xtag, and
FrameNet




Summary of the theoretical work done on
semantic-verb classification for the decades
leading up to it
Primary verb classes are based upon
alternation
Also dependent on extended meaning of
verbs and morphology
Taxonomy which provides a classification of
3,024 verbs (4,186 senses)

Puts these verbs into 48 broad and 192 finegrained classes according to the 79
alternations with NP and PP complements



“Break verbs” are verbs which bring about a
change on the material integrity of an entity
Characterized by its participation or nonparticipation in alternations and other
constructions
8 different examples



Causative/inchoative alternation:
Tony broke the window <-> The window
broke
Middle alternation
Tony broke the window <-> The window
broke easily
Instrument subject alternation
Tony broke the window with the hammer <->
The hammer broke the window



With/against alternation
Tony broke the cup against the wall <-> Tony
broke the wall with the cup
Conative alternation
Tony broke the window <-> Tony broke at the
window
Body-Part possessor ascension alternation
Tony broke herself on the arm <-> Tony broke
her arm


Unintentional interpretation available
Reflexive object: Tony broke himself
Body-part object: Tony broke his finger
Resultative phrase
Tony broke the piggy bank open <-> Tony
broke the glass to pieces




Hierarchical domain-independent
Higher level of detail then the original
classification of the word in things such as
WordNet
Includes both syntactic and semantic
information for classes of English verbs based
of Levin’s classification
Much more detail than is given in the original
classification such as what WordNet has


Verb classes, like the “Break verbs” example,
are described by the set of members,
argument structure, restrictions, and frames
of the semantic and syntactic predicates
This type of structure allows for a varying
level of granularity which can be changed
based on the type of NLP application being
used



Syntactic frames consist of thematic roles
(such as agent, theme and location), the verb,
and any other lexical items needed for
alternation or construction of the verb
May also be limited by which prepositions are
allowed
Semantic restrictions (such as animate,
human, and organization) are used to
suggest the preference of thematic roles
allowed by the classes



The first version of VerbNet had 4,100 verb
senses (over 3,000 lemmas) distributed
across 191 first-level classes, and 74
subclasses
These desciptions used 21 thematic roles, 36
selectional restriction preferences, 314
syntactic frames and 64 semantic predicates
Also dependent on a hierarchy of 57
prepositions


The first version of VerbNet was evaluated
through a mapping to roughly 50,000
instances of PropBank (lexicon of
propositions and their arguments) corpus
instances
The syntactic frames for VerbNet accounted
for over 78% of exact matches found in the
frames of PropBank



In 2006 VerbNet was expanded and from
Korhoren and Briscoe (2004) there were 57
new novel classes added to the original Levin
classification which VN was based
Also added 106 new diathesis alternations
This added to the coverage of PropBank and
boosted it from the once 78% all the way up
to 90.86%



Electronic resource based on a theory called
frame semantics (getting the same message
across from a different perspective)
Lexical database which contains 1,200
semantic frames, 13,000 lexical units (pairing
of a word with a meaning) and over 190,000
example sentences
Mainly created by Charles J. Fillmore, the
creator of frame semantics and intial project
leader in 1997 when the project began

VerbNet:
 Parameterized Action Representation (work done
at Penn which bridges the gap between natural
language and animation)

FrameNet:
 Question answering, paraphrasing, textual
entailment and information extraction

VerbNet and FrameNet
 Unified Verb Index
Consists of VerbNet, PropBank ,FrameNet and
OntoNotes Sense Groupings
 Merges links and webpages from 4 different
natural language processing systems
 8, 537 total verbs represented
 6, 340 VerbNet links
 273 VerbNet main classes
 214 VerbNet subclasses
 5, 649 PropBank links
 4, 186 FrameNet links
 4, 898 total Grouping Links






VerbNet
Levin Classification
FrameNet
Applications
Unified Verb Index
http://bubblegls.com/2012/02/05/verbnet-and-framenetlexical-semantics-iii/
 http://verbs.colorado.edu/~kipper/Papers/lrec-journal.pdf
 http://verbs.colorado.edu/~mpalmer/projects/verbnet.html
 http://en.wikipedia.org/wiki/VerbNet
 http://faculty.ist.unomaha.edu/ylierler/teaching/material/fra
menet.pdf
 https://framenet.icsi.berkeley.edu/fndrupal/
 http://en.wikipedia.org/wiki/FrameNet
