I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
Linking WordNet Verb Classes to Semantic Interpretation 
Fernando Gomez 
School of Computer Science, University of Central Florida 
Orlando, Fl 32816 
(gomez~cs.ucf.edu) 
Abstract 
An analysis that defines predicates for Wordnet verb 
classes and links them to semantic interpretation is 
presented. The selectional restrictions for the the- 
matic roles defining the predicates are WordNet on- 
tological categories. Thematic roles are also linked 
to the syntactic relations that realize them. The pa- 
per illustrates the methodology by providing a de- 
tailed analysis of some major WordNet verb classes. 
1 Introduction 
We present an analysis i of some WordNet verb 
classes (Miller et al., 1993; Fellbaum, 1993). Its 
purpose is to offer a detailed analysis of some verb 
classes so that the reader may grasp the main ideas 
guiding our methodology for constructing verb pred- 
icates using the WordNet lexicon. We plan to pub- 
lish a complete listing of all our predicates and sub- 
predicates for all WordNet verb classes, once they 
have been fully tested. In (Gomez, 1998), the reader 
can find a detailed description and evaluation of the 
semantic interpreter algorithm that uses the lexi- 
cal entries defined here. This paper complements 
(Gomez, 1998) in which, for space reasons, only one 
predicate could be analyzed. 
The hierarchical decomposition of predicates plays 
a central role in our methodology. A generic predi- 
cate subsumes all its subpredicates in a similar way 
in which the generic concept "book" subsumes all its 
subconcepts. A subpredicate inherits thematic roles 
and inferences from its super-predicate. From the 
point of view of the lexicographer, inheritance comes 
with an added bonus because it allows her/him to 
handle a large number of verbs by just defaulting 
their lexical entry to its super-predicate. In a sub- 
sequent analysis, the lexical entries can be refined 
by mapping some verbs which were just defaulted to 
a superpredicate into their own subpredicates. This 
top down approach and the WordNet lexicon makes 
possible to define predicates for every English verb 
in a "reasonable" amount of time. The mapping of 
Q 
XThis paper contains some material published aa report 
UCF-CS-TR-97-03 January-4-1997 with the same title. 
58 
WordNet verb classes into generic predicates has re- 
quired to define new classes and to reclassify and/or 
redefine some WordNet classes and subclasses. The 
WordNet ontology for nouns also has been reorga- 
nized and redefined in some respects in order to con- 
form with the entries in the thematic roles of the 
predicates. (See (Gomez, 1998) for some details.) 
Two major consequences derive from anchoring verb 
classes in abstract semantic predicates: coalescing 
several WordNet synsets into a predicate, and map- 
ping the same WordNet synset into distinct pred- 
icates. The differentia between a generic predicate 
and its subpredicates are given by one or more of the 
following: a) specific selectional restrictions for the 
thematic roles, b) different syntactic realizations of 
the thematic roles, and c) specific sets of inferences 
associated with the subpredicates. 
The semantic interpreter algorithm, which is an 
extension of the one reported in (Gomez et al., 1997), 
is based on the idea that the meaning of the verb de- 
pends not only on its selectional restrictions, but also 
on the syntactic relations that realize them. A sim- 
ilar view has been presented in (Pustejovsky, 1995). 
Semantic interpretation is delayed until the end of 
a clause. For every verb in a clause, WordNet pro- 
vides a list of verb synsets for which we have defined 
predicates. These predicates are contenders for the 
meaning of the verb. For every syntactic relation in 
the clause, the interpreter checks each predicate in 
order to see if the predicate has a thematic role which 
is realized by the syntactic relation. The interpreter 
records this fact and gets the next syntactic rela- 
tion. The predicate that realizes the most syntactic 
relations in the sentence is selected as the meaning 
of the verb. In case of ties, the predicate that has 
more thematic roles realized is selected. 
2 The syntax of roles 
The syntax of roles in the predicates that will be de- 
fined below is ~ven by the cfg grammar on the next 
page. Each thematic role is followed by any number 
of list pairs. The first list contains the selectional 
restrictions, a subset of the ontological categories in 
WordNet, in order of preference (Will<s, 1975) for 
! 
! 
the thematic role, and the second list contains the 
syntactic relations (henceforth, SRs) that may real- 
ize the thematic role. For any given sentence, only 
one of the SIRs in the list realizes the role. An onto- 
logical category preceded by the sign %" in a selec- 
tional restriction means that any noun sense that is 
subsumed by the semantic category preceded by "-" 
does not realize that thematic role. 
S -> (ROLE L); ROLE -> agentJtheme\[.. 
L-> (SM) (SR) L\[ (SM) (SR) 
SM -> 0NTOLOGICAL-CAT SM I 0NTOLOGICAL-CAT 
SM -> -0NTOLOGICAL-CAT SM 1-0NTOLOGICAL-CAT 
SR -> SYNTACTIC-REL SR I SYNTACTIC-REL 
SYNTACTIC-REL -> (prep PREP) I subj ~ 
\[ obj \[obj2\[ subj-if-obj \[ subj-if-no-obj f 
\[ obj-if-obj 2 \[ predicate-complement \[ 
\[ complement-phrase } adj ect ire-phrase 
PREP -> ANY-PREP PREP J ANY-PREP 
0NTOLOGICAL-CAT ->thing \[ physical-thing f... 
ANY-PREP -> on \[ in \[ .... 
The entry obj refers to the first postverbal NP, obj2 
to the second postverbal NP. Subj-if-obj refers to the 
subject of a sentence that also has an object (the 
verb is used transitively), and subj-if-no-obj refers 
to the subject of a sentence containing no object 
(the verb is used intransitively). Obj-if-obj2 refers 
to the obj of a sentence having also an obj2 (the 
verb is used ditransitively). Thus, subj refers to the 
subject of a sentence without expressing any context 
about the transitivity or intransitivity of the verb. 
3 Verbs in which the agent causes a 
change of location of something 
else 
In (Gomez, 1998), we provide an analysis of verbs in 
which an animate agent changes location. We start 
explaining the predicate cause-to-change-location 
depicted on top of the next column. The primary 
event expressed by this predicate is a cause of change 
of location of something other than the agent; al- 
though the agent may have also changed location. 
In "Kelly carried the flowers to the table" and in 
"Kelly drove John to school," the agent has also been 
moved, but the primary event is the fact that Kelly 
causes a change of location of the flowers and John, 
respectively. The WordNet synset "move2, displace, 
make move - (cause to move)" is coalesced, in prin- 
ciple, into this predicate. But, also many of the sub- 
classes of "move3, change position" are mapped into 
subpredicates of this predicate. (Physical-thing has 
been abbreviated to phy-thing in some entries for 
space reasons.) The urn-map entry means that all 
verb forms under the synset move2 are in principle 
coalesced into the concept cause-to-change-location. 
However, subpredicates of this predicate will be rec- 
ognized as explained below. The category human- 
59 
agent subsumes human and social-group. The role 
inanimate-cause stands for an agent that is not an 
animate being or an organization. The syntactic re- 
alization of this role is always a sub j-if-obj. Because 
the agent of this predicate is also realized by subj-if- 
obj, and the ontological category animal and human- 
agent are subsumed by physical-thing, then in "Beth 
put the books on the table," both the agent and the 
inanimate-cause will match. The impasse is solved 
by preferring the agent over the inanimate-cause. 
The theme is realized by obj, and is also realized by a 
subj if the verb is used intransitively. Thus, for "The 
wind has knocked down many apples from the tree," 
"the wind" is the inanimate-cause, "many apples" is 
the theme and "from the tree" is the source. In "The 
leaves have moved," the "leaves" is the theme, and 
the agent, or inanimate-cause is unknown. In "The 
tractor has pulled the tree" "the tractor" is the in- 
strument, and the agent is unknown. The different 
ways of expressing a path, namely toward-goal, near- 
goal, etc. have been collapsed here into goal for space 
reasons. Actually, there is one role for each preposi- 
tion that expresses a different spatial relationship. 
\[cause-to-change-location(is-a (action)) 
(~m-map(move2)) 
(agent(human-agent animal)(subj-if-obj)) 
(theme(phy-thing) (obj subj-if-no-obj)) 
(source(location phy-thing) ((prep from))) 
(goal(location phy-thing)((prep to tovaxds 
toward in through into back-to along over 
beside above by on under below throughout 
beyond past across near up))) 
(instz~tment(instrumental£ty animal-body-part) 
(subj-if-obj((prep with on in))) 
(animal) ((prep on))) 
(distance(distance linear-measure) 
((prep for))) 
(inanimate-cause(phenomenon physical-thing) 
(subj-if-obj))\] 
The first subclasses of move2 analyzed by Word- 
Net are the senses of "transport." The subclasses 
formed by the synsets "transportl (move something 
or somebody around; usually over long distances)," 
"transport2, carry (move while supporting, either 
in a vehicle or in one's hands or on one's body)," 
"transport3, send, ship - (transport commercially)" 
and "transmit, transfer, conduct, transport5, chan- 
nel, channelize" are mapped into subpredicates of 
the predicate transport which is: 
\[transport(is-a(cause-to-change-location)) 
(theme (physical-tbing)(obj obj2)) 
(goal(human-agent animal) 
(obj-if-obj2 (prep for))) 
(location phy-thing)((prep to towards ... 
same as cause-to-change-location)))\] 
\[transport -over-long-distances 
(is-a(transport)) (wn-map(transpor~l))\] 
<other transport predicates here> 
The goal, besides being realized by the same prepo- 
sitions as those for cause-to-change-location, is also 
syntactically realized by an indirect object (obj-if- 
obj2), e.g., (1)"Susan brought her children a book 
from Harvard." Thus, the theme can be realized by 
an obj2 or by an obj, e.g., "Mary brought her chil- 
dren from school." This is the analysis that this 
work provides for all double object verbs. Because 
the semantic interpreter delays commitment until 
the end of the clause, the interpreter does not have 
to reanalyze. 
The predicate pull, which corresponds to the 
synset "pulll, draw, force," and that contains such 
forms as "jerk," "twitch," "trail," "drag," etc. is: 
\[pull (is-a (cause-to-change-location)) 
(vn-map (pulll)) 
(agent (human-agent animal) (subj-if-obj)) 
(theme(instrumentality physical-thing) 
(obj (prep from off))) 
(source(location phy-thing) ((prep from 
off)))\] 
Next, WordNet includes a large class of forms un- 
der the synset "put, set, place, pose, position, lay; 
also with abstract objects and locations." This work 
differs from WordNet and does not include abstract 
objects as themes of this predicate. The predicate 
put is: 
\[put (is-a (cause-to-change-location)) 
(~ra-map (purl)) 
(theme (physical-thing) (obj)) 
(goal(location physical-thing) ((prep on 
in towaxds through into back-to along 
over beside above by under below 
throughout beyond past across near))) 
(instrument (instrumentality) ((prep with))) 
(source (nil) (nil))\] 
This predicate illustrates two points. First, its goal is 
not expressed with the preposition '%o," e.g., *Mary 
placed/arranged/put/etc the books to the table. Sec- 
ond, the source is not realized by any syntactic re- 
lation (Levin, 1993). That is why the source entry 
is nil for the selectional restriction and the syntac- 
tic relation in order to indicate that this predicate 
does not inherit a goal from its ancestors. The con- 
cern here is not syntactic subcategorization of the 
verb forms, which is handled by the parser, but at- 
tachment of PPs. Any PP whose preposition ap- 
pears in th~ entry of a thematic role is going to 
be attached preferentially to the verb. Thus, in 
"Connie put the reports from MIT on the shelf," 
60 
the prepositional attachment algorithm should not 
consider "from MIT" as a possible complement of 
"put." WordNet includes several subclasses of putl, 
all of which have been mapped to subpredicates of 
put. Some verb forms under this predicate, such 
as "put" and "place," realize their goal with ~at" 
phrases. This fact is represented in the verb forms 
themselves. The next major subclass of move2 is 
under the synset "propell, impel, move forward," 
which includes the subclasses "hitl cause to move by 
striking") and "throw" verbs. The predicate propel 
is depicted on the next column. Most of its thematic 
roles are inherited from cause-to-change-location, il- 
lustrating how relatively easy it is to build these en- 
tries once the predicates for the main subclasses have 
been constructed. The goal of this predicate is also 
realized by an indirect object (obj-if-obj2), e.g., ~Sue 
threw the pitcher a ball." This class of verbs has 
an argument realized by the preposition "against," 
that has been mapped into the role contact-goal, a 
refinement of the role goal, meaning that the theme 
contacted the goal 
\[propel(is-a (cause-to-change-location)) 
(ira-map (propel1)) 
(theme (physical-thing) (obj obj2)) 
(goal(human-agent) (obj-if-obj2) 
(location physical-thing) 
((prep to on in through towards 
into back-to along over beyond past 
across by near at))) 
(contact-goal(physical-thing)(against))\] 
The subclass formed by hitl, not to be con- 
fused with "hit3 - ideal a blow to; He hit her 
hard in the face)," becomes \[cause.to-move-by-biting 
(is-a(propel))(wn-map(hitl))\], the one formed by 
"throwl, project through the air" becomes \[throuJ(is- 
a(propel))(urn-map(throwl))\]. A subclass of throwl 
is formed by those verbs that Levin (Levin, 1993) 
calls ~pelt" verbs ("buffet," "bombard," "pelt," 
"shower," "stone,) in which the goal is realized by 
obj and the theme by a "with" phrase, e.g., "Beth 
pelted Chris with snowballs." Of these, WordNet 
does not include "shower" and "stone" as forms un- 
der peltl. 
Some classes in move2 and move3 are not mapped 
into subpredicates of the generic predicate cause- 
to-change-location. The reasons for not mapping a 
WordNet subclass into a subpredicate of the generic 
predicate for the WordNet class are any one of the 
following: a) these predicates do not share the the- 
matic roles of the generic predicate, b) the primary 
event expressed by these subpredicates is not that of 
the generic predicate, or c) the ontological categories 
in the selectional restrictions of the generic predicate 
do not subsume those in the subpredicates. One 
major subclass of move2 that clearly is not a sub- 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
predicate of cause-to-change-location is grouped un- 
der the synset "transfer, convey, transmit, communi- 
cate, bring," and includes such forms as "relay," "re- 
quest," "demand," "expect," "require," "beg," and 
many others. The theme of the predicate for this 
class is not a physical-thing, but an abstract thing, 
idea or a proposition, which is many times realized 
by a sentential complement. Two other strongly 
related subclasses of move2 that are not subpredi- 
cares of cause-to-change.location are grouped under 
the synsets "expel, eject, chuck out, exclude, throw 
out, kick out, turf out, boot out, turn out - (from a 
place)," and "expel, throw out, kick out - (as from a 
country)." The first subclass includes such forms as 
"evict," "exorcise," and the second subclass includes 
"expatriate," "deport," "disbar," "extradite," "os- 
tracize," and others. The relation between the theme 
and the agent in the predicates underlying most of 
these forms is an abstract relation rather than a 
physical relation. The analysis provided for these 
classes has been to create a subpredicate of action, 
called expel-human-agent, which coalesces the two 
synsets discussed above plus a synset of remove, re- 
move2, a class by itself, "remove - (remove from a 
position or an office)." Remove2 includes such forms 
as "fire," "lay off," etc. This generic predicate is an- 
alyzed into several subpredicates. 
Another major subclass of move2 that is not 
mapped into a subpredicate of cause-to-change- 
location is grouped under the synset "separate2, dis- 
unite, force apart, divide, part, take apart, pull apart 
- He separated the fighting children; Moses parted 
the Red Sea," and includes such forms as "com- 
partmentalize," "polarize," "disarticulate," "discon- 
nect," "cut" and its subclasses, and many oth- 
ers. The primary event of this subclass is not a 
cause-to-change-location, although one may be im- 
plied. Some of these forms can be used in the 
causative/inchoative alternation, e.g., "The cream 
separated from the milk," and in the middle al- 
ternation, e.g., "Cream separates easily from milk" 
(Levin, 1993). Separate2 is analyzed as a subpred- 
icate of cause-a-change-of-state, and coalesced with 
some other synsets of "separate" which are classi- 
fied by WordNet as subclasses of change1, cause a 
change of state. Another subclass of move2 that has 
required a special analysis is that of "chase away, 
drive out, turn back, drive away, dispel, drive off, 
run off" in which the agent is causing the theme, 
an animate being, to move away from the agent's 
location. 
In move3, some individual forms and some sub- 
classes are mapped into the predicate move-body- 
part, a subpredicate of action. Some of these sub- 
classes are: "move involuntarily, move reflexively - 
(move in an uncontrolled manner)," "dance, do a 
dance ..." and a few others. 
61 
The following predicates do not come from sub- 
classes of move2 or move3, but from change1 (cause a 
change of state). However, they are analyzed as sub- 
classes of cause.to-change-location. The first predi- 
cate fill-or-load, depicted on the next column, co- 
alesces the synsets: "fill1, fill up, make full," and 
"flU2, fill up, become full." Of these, flU1, which is 
a subclass of change1 contains most of the forms, 
including all the "load" verbs. Fill2, a subclass 
of change2 (undergo a change) contains two forms. 
Note the different syntactic realizations of this predi- 
cate, as exemplified by "The farmer loaded the truck 
with hay" and "John loaded the tractors into the 
ship." It is not very helpful to include inntrumental- 
ity as one of the selectional restrictions of the theme 
because anything can be loaded. However, substance 
could help to choose the correct sense of certain 
nouns, because it is a selectional restriction used 
frequently with "fill" verbs. The predicate empty 
that coalesces the synsets "empty1, make empty, 
make void" and "empty2, become empty, become 
void, discharge" becomes a subpredicate of remove- 
physical-thing, described below. The obj-if-with re- 
lation in the goal role means that this role can be 
realized by an obj if there is a with-phrase. For in- 
stance, "Kelly loaded the truck with hay." 
If ill-or-load (is-a (cause-to-change-location)) 
(wn-map (filll) (fill2)) 
(theme(substance physical-thing) (obj 
( (prep with) ) ) ) 
(goal(instrumentality physical-thing) (obj 
obj-if-with (prep into on onto in)))\] 
The next major class is grouped under the synset 
of "remove1, take, take away - (remove something 
concrete, as by lifting, pushing, taking off, etc.; or 
remove something abstract; remove a threat ...)." 
Remove2 was analyzed above in the predicate expel- 
human. Remove1 forms a class by itself in WordNet. 
\[remove-physical-thing (wn-map (remove1)) 
(is-a (cause-~o-change-loca~ion)) 
(theme(physical-thing) (obj)) 
(source(physical-thing) ((prep off from))) 
(goal(nil) (nil))\] 
Removel contains many subclasses, most of which 
are mapped to subpredicates of remove-physical- 
thing. Remove3: "get rid of, remove," also forms 
a class by itself and is analyzed as a subpredicate 
of remove-physical-thing. WordNet does not include 
"rid" as a subordinate of any of the "remove" senses, 
but as a small class by itself containing nine forms. 
This work maps "rid1, free, disembarrass" to r/d, 
a subpredicate of remove-physical-thing, in which 
the theme is realized by an "of" phrase, and the 
source by obj, e.g., "He rid the city of rats." Empty, 
another subpredicate of remove-physical-thing, may 
I 
I 
I 
also realize its theme and goal with an "of" phrase 
and obj, respectively. Some forms under this sub- 
class are used in the sense of "firing somebody." 
Rather than to reclassify these forms, a meaning 
postulate connects this predicate to the predicate 
expel-human-agent, explained above, if its theme is 
the concept human. This subclass also contains a 
subclass that does not express a source, but a goal. 
This subclass is: "abandonl - (We abandoned the 
old car in the empty parking lot)," which includes 
such forms as "chuck," "ditch," "dispense with," 
and others. AbandonS: "abandon, forsake, desolate, 
desert, lurch - (leave someone who needs or counts 
on you);" which is a suclass of "leave2" is also coa- 
lesced into the the predicate: \[abandon-phy-thing- 
or-animate (is-a(get-rid-of ) ) (wn-map(abandonl ) O veZ)) (sou e(nit)(nit))\] 
4 Transfer of Possession Verbs 
Table 1 depicts the hierarchy of subpredicates of 
transfer-of-possession. 
sell lease 
pay invest 
trade spend-physical-thing 
import use-up 
export squander 
give get-something 
give-back capture-an-animal 
bestow receive-something 
feed graduate 
nourish buy 
bribe subscribe 
distribute accept-something 
deposit-money accept-a-job 
lend accept-somebody 
offer-something obtain 
provide get-back 
serve-somebody reclaim 
arm win-at-an-event 
fuel-an-engine score-in-a-game 
equip win-in-an-election 
shelter gain-something 
donate profit 
bequeath inherit 
endow borrow 
gather-things 
get-by-force-ilegly 
~nnez 
steal 
Table 1: Hierarchy for subpredicates of transfer of possession 
The major subpredicates are give and get which 
are listed last. The generic predicate transfer-of- 
possession is described below. There are two events 
in a transfer of possession. The theme refers to the 
62 
thing obtained or received by the agent of the pri- 
mary event, and the co-theme to the thing obtained 
or received by the agent of the secondary event. A 
meaning postulate infers the predicate and its roles 
for the secondary event. The exclusionary semantic 
categories in the co-theme are to impede indetifying 
a \[for NP\] as a co-theme when it is a to-poss, e.g., 
"He bought a book for Mary/20 dollars." 
\[transfer-of -possession 
(is-a(action)) 
(wn-map (trans f er6) ) 
(agent (human-agent animal) (subj)) 
(theme(possession thing) (obj obj2)) 
(to-poss (human-agent animal) 
(obj-if-obj2 (prep for to)) 
(physical-thing) ((prep to))) 
(c o-theme (po s s e ss ion -human-agent -animal 
physical-thing thing) ( (prep for))) \] 
The subpredicate give is depicted below. The only 
difference between give and its generic predicate is 
that it does not have a co-theme. The synset give3 
has many verb forms. Some of them are analyzed 
below. 
\[give 
(is-a (transf er-of-possession) ) 
(wn-map (give3)) 
(co-theme(nil) (nil))\] 
The predicate feed depicted below has some inter- 
esting garden path constructions. Compare "Beth 
fed the frogs insects" to "Ann fed the frogs to the 
snakes" and "Kathy fed the frogs." 
freed (is-a(give)) 
(wn-map (feedl) (feed2)) 
(agent (animal) (subj-if-obj)) 
(theme(food substance physical-thing) 
(obj-if-to obj2) ) 
(to-poss (animal human-agent) 
(obj obj-if-obj2 (prep to)))\] 
The synsets supply1 and provide2 are mapped 
into two predicates, provide and provide-inanimate- 
cause depicted on the next page. The predicate 
provide-inanimate-cause subsumes all those predi- 
cates in which the thing that causes the transfer 
is not an animate being, but an inanimate thing, 
e.g., "The river provides water to many cities." The 
syntactic relation obj.if-with indicates that the role 
to-poss can be realized by an obj if this is followed 
by a \[with NP\]. This is necessary in order to handle 
the sentence "France also provided the missionary 
with new churches." The next mayor subpredicate of 
transfer-of-possession is get-something depicted be- 
low. Its synset, getl, is a unique class in WordNet. 
One of the major subclasses of get-something is the 
predicate capture-an-animal which has a very simple 
entry. 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
\[provide 
(is-a(give)) 
(ran-map (supply1) (provide2)) 
(theme (t:hing) (obj obj 2 (prep with))) 
(to-poss (human-agent animal 
physical-thing thing) 
(obj-if-obj2 obj-if-with 
(prep to for))) 
(inanimate-cause (thing) (subj -if-obj ) ) \] 
\[provide-inanimate-cause 
(is-a(give)) 
(agent (nil) (nil)) 
(wn-map (supply 1) (provide2)) 
(theme (physical-thing) (obj obj 2 (prep with)) ) 
(to-poss (human-agent animal 
physical-thing thing) 
(obj-if-obj2 obj-if-with 
(prep to for))) 
(inanimate-caus e (-human-agent -animal 
thing) (subj-if-obj)) \] 
\[get-something 
(is-a(transf er-of-possession) ) 
(un-map (get I) ) 
(theme(physical-thing thing) 
(obj obj2)) 
(from-poss (human-agent animal phy-thing 
thing) ((prep off from out-of))) 
Most of its subpredicates require a human as agent, 
but some take an animal as agent. Another class is 
formed by the verbs belonging to the synset receivel 
and receive2 whose predicate is: 
\[receive-something (is-a (get-something)) 
(un-mapCreceivel) (receive2)) 
(theme(auard -human-agent physical-thing 
-perception thing) (obj obj2))\] 
The exclusionary categories human-agent and per- 
ception are intended to exclude the sense of receiv- 
ing, or welcoming somebody, and that of receiving 
experiences, or experiencing something, from this 
predicate. Those senses are mappped into their own 
predicates. 
The synset winl (be the winner in a contest or 
competition) that is a unique class in WordNet 
has been mapped into the predicate win-at-an-event 
which has become a subconcept of get-something. 
The synset win2 (win something) is mapped into 
the predicate gain-something. Two other subpred- 
icates of get-something whose synsets form unique 
classes in WordNet are gather-things (gatherl) and 
get-by-/orce-or-illegally (takel4). 
5 Conclusions and Related Research 
We have described a method for defining predicates 
for WordNet verb classes and illustrated it by ann- 
63 
lyzing some major WordNet verb classes. The the- 
matic roles of the predicates are connected to Word- 
Net ontology for nouns and to the syntactic rela- 
tions that realize them. Our approach to building 
verb meaning is based on the decomposition of a 
predicate into subpredicates. Subpredicates inherit 
not only thematic roles, but also inferences as ex- 
plained in (Gomez, 1996). Inferences will be defined 
for generic predicates and subpredicates subsuming 
a large class of verb forms. The final product will 
be a network of predicates linked by subsumption, 
and inferences, and connected to the WorcbNet on- 
tology for nouns and to grammatical relations. As 
of this writing we have defined predicates for 70% of 
WordNet verb forms. 
We are using the term "inference" to refer to 
both entailments and non-monotonic inferences, also 
called defeasible inferences. The term "entailment" 
is used in the sense of analytic implication (Quine, 
1960). A sentence, say sl, entails sentence s2 if s2 is 
true in every world in which sl is true. An example 
of entailment is "if Peter killed the roach, the roach 
is dead." An example of non-monotonic inference is 
"if Peter likes apples, he eats them." Of course, it is 
very important to bear in mind that the difference 
between non-monotonic inferences and entailments 
is a question of degrees as Quine (Quine, 1953) has 
argued convincely. See (Gomez, 1996) for a discus- 
sion. 
Our work differs from the semantic role list ap- 
proaches (Fillmore, 1968) in several essential as- 
pects. First of all in our method, the semantic roles 
are not defined independently of the meaning of the 
verb and are not semantically unanalyzable. In addi- 
tion, the number of thematic roles depends on each 
predicate, and not on some general criteria saying 
which thematic roles there will be, irrespective of 
each predicate. Any thematic role in a predicate 
corresponds to a meaning relation from which in- 
ferences specific to that predicate or subpredicates 
must be established. Consider the sentence "These 
birds fly 11,000 miles from their breeding grounds 
to their winter home in the Antarctic." What is 
the thematic role corresponding to the NP "11,000 
miles?" Some semantic relation needs to be rec- 
ognized so that inferences such as "there is a dis- 
tance of 11,000 miles between the breeding grounds 
of these birds and their winter home" can be es- 
tablished. We have recognized that semantic rela- 
tion by creating the role distance meaning the dis- 
tance traveled by an animate agent in a change-of- 
location predicate (Gomez, 1998). The inference is 
based on that role and on that predicate. Because 
the NP "11,000 miles" is not a subcategorized argu- 
ment of "fly," why call it a "role" and not an "ad- 
junct"? From a semantic point of view, it makes 
no difference whether one calls it "adjunct" or "the- 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
matic role." Dowry (Dowty, 1991), then, asks that, 
if one assigns a thematic role to measures of dis- 
tance, why not assign a role to "quickly" in "She 
walks quickly." Our answer is that we should as- 
sign a role to "quickly." But, that role, whatever its 
name, should not be placed in the predicate change- 
o\]-location but in the action node because it can be 
inherited by every subpredicate of action. It makes 
sense to say "She eats/studies/writes/.., quickly," 
but not "She eats 20 miles." Thus from our point 
of view, an "adjunct" is a role that is inherited by 
every subpredicnte of action. 
This approach does not lead to a propagation of 
roles since their number and nature depend on the 
generic predicate and its subpredicates. The critique 
of "role fragmentation - the subdivision of a single 
role into many subroles as result of subsequent anal- 
ysis (Dowry, 1991) - is valid ff the entailments are 
based exclusively on the role, but not if they are an- 
chored on the role and the predicate. The roles that 
we have used throughout our analysis have differ- 
ences in meaning across diverse generic predicates, 
or verb classes. For instance, the meaning of the 
role theme in a change of state verb, say "break," 
is different from its meaning in a transfer of infor- 
mation verb, say "tell." Hence, if the entailments 
are based only on the role, one would be compelled 
to recognize several types of theme (Dowty, 1991), 
but because the entailments are based on the pred- 
icate and on the role, this is not necessary. Role 
entailments are shared by subpredicates of a generic 
predicate not across generic predicates. 
Our approach also differs from those analyses that 
attempt to reduce the verb semantic analysis to a 
small set of notions e.g., Jackendoff'localist hypoth- 
esis (Jackendoff, 1990), Dowty-Vendler's aspectual 
hypothesis (Vendler, 1967), Dowty (Dowry, 1979), 
or to a small set of primitives (Schank, 1975). Our 
major critique to reductionist analyses are Quinean 
(Quine, 1960) in nature, namely meaning is holis- 
tic. Trying to reduce verb meaning to a small set of 
notions is going to fail because verb meaning is inter- 
sentential. One may take some verbs whose meaning 
can be reduced to some few principles, or notions. 
Verbs of change of location fit very well within the 
localist hypothesis, while verbs of creation seem to 
support the aspectual hypothesis. But, if you con- 
sider a verb like "graduate" the situation is rather 
different because many sentences mediate to form 
its meaning. One may say that it is an accomplish- 
ment verb, but that is not saying much. Trying to 
provide a CD representation (Schank, 1975) for it 
seems hopeless. The key point here is that the repre- 
sentation needs to make possible the inferences nor- 
madly associated with such verb. For instance, that 
if one is asked which school Clinton attended, one 
should say "Yale" upon reading that Clinton grad- 
64 
uated from Yale. Hierarchical predicate decomposi- 
tion and inferences shared by subpredicates of the 
same generic predicate is the solution that we offer. 
However, we always keep in mind that meaning is 
a question of degrees. It can be approximated, but 
not fully grasped. Fortunately, approximation is all 
we need for the task of natural language processing. 

References 
D. R. Dowty. 1979. Word Meaning and Montague 
Grammar. Dordrecht: Reidel, Dordreeht. 
D Dow-ty. 1991. Thematic proto-roles and argument 
selection. Language, 67:547-619. 
C. Fellbaum. 1993. English verbs as a semantic net. 
Technical report, Princeton. CSL Report 43, re- 
vised March 1993. 
C. J. Fillmore. 1968. The case for case. In E. Bach 
and R.T. Harms, editors, Universals in Linguistic 
Theory, pages 1-88. Holt, Rinehart and Winston, 
New York, NY. 
F. Gomez, C. Segami, and R. Hull. 1997. Determin- 
ing prepositional attachment, prepositional mean- 
ing, verb meaning and thematic roles. Computa- 
tional Intelligence, 13(1):1-31. 
F. Gomez. 1996. Acquiring intersentential ex- 
planatory connections in expository texts. In- 
ternational Journal o\] Human-Computer Studies, 
4(1):19-44. 
F. Gomez. 1998. Semantic interpretation of un- 
restricted domains using wordnet. CS-TR-98-2, 
Jan. 1998. 
R. Jackendoff. 1990. Semantic Structures. MIT 
Press, Cambridge, Mass. 
B. Levin. 1993. English Verb Classes and Alterna- 
tions: A Preliminary Investigation. University of 
Chicago Press, Chicago. 
G.A. Miller, R. Beckwith, C. Fellbaum, D. Gross, 
and K. Miller. 1993. Introduction to WordNet: 
An on-line lexical database. Technical report, 
Princeton. CSL Report 43, revised March 1993. 
J. Pustejovsky. 1995. The Generative Lexico~ MIT 
Press, Cambridge, Mass. 
V.W. Quine. 1953. Two dogmas of empiricism. In 
From a Logical Point o/View. Harvard University 
Press, Cambridge, Mass. 
V.W. Quine. 1960. Word and Object. MIT Press, 
Cambridge, Mass. 
R. Schank. 1975. Conceptual Information Process- 
ing. North Holland, Amsterdam. 
Z. Vendler. 1967. Linguistics in Philosophy. CorneU 
University Press, Ithaca, New York. 
Y.A. Wilks. 1975. Preference semantics. In E.L. 
Keenan, editor, Formal Semantics of Natural Lan- 
guage. Cambridge University Press, Cambridge, 
UK. 
