Investigating regular sense extensions based on intersective Levin 
classes 
Hoa Trang Dang, Karin Kipper, Martha Palmer, Joseph Rosenzweig 
Department of Computer and Information Sciences 
and the Institute for Research in Cognitive Science 
University of Pennsylvania 
400A, 3401 Walnut Street/6228 
Philadelphia, PA 19104, USA 
htd/kipper/mpalmer/josephr@cis.upenn.edu 
Abstract 
In this paper we specifically address questions of 
polysemy with respect to verbs, and how regular 
extensions of meaning can be achieved through 
the adjunction of particular syntactic phrases. 
We see verb classes as the key to making gen- 
eralizations about regular extensions of mean- 
ing. Current approaches to English classifica- 
tion, Levin classes and WordNet, have limita- 
tions in their applicability that impede their 
utility as general classification schemes. We 
present a refinement of Levin classes, intersec- 
tive sets, which are a more fine-grained clas- 
sification and have more coherent sets of syn- 
tactic frames and associated semantic compo- 
nents. We have preliminary indications that 
the membership of our intersective sets will be 
more compatible with WordNet than the orig- 
inal Levin classes. We also have begun to ex- 
amine related classes in Portuguese, and find 
that these verbs demonstrate similarly coherent 
syntactic and semantic properties. 
1 Introduction 
The difficulty of achieving adequate hand- 
crafted semantic representations has limited the 
field of natural language processing to applica- 
tions that can be contained within well-defined 
subdomains. The only escape from this lim- 
itation will be through the use of automated 
or semi-automated methods of lexical acquisi- 
tion. However, the field has yet to develop a 
clear consensus on guidelines for a computa- 
tional lexicon that could provide a springboard 
for such methods, although attempts are being 
made (Pustejovsky, 1991), (Copestake and San- 
filippo, 1993), (Lowe et al., 1997), (Dorr, 1997). 
The authors would like to acknowledge the sup- 
port of DARPA grant N66001-94C-6043, ARO grant 
DAAH04-94G-0426, and CAPES grant 0914/95-2. 
One of the most controversial areas has to do 
with polysemy. What constitutes a clear sepa- 
ration into senses for any one verb, and how can 
these senses be computationally characterized 
and distinguished? The answer to this question 
is the key to breaking the bottleneck of semantic 
representation that is currently the single great- 
est limitation on the general application of nat- 
ural language processing techniques. 
In this paper we specifically address questions 
of polysemy with respect to verbs, and how 
regular extensions of meaning can be achieved 
through the adjunction of particular syntactic 
phrases. We base these regular extensions on 
a fine-grained variation on Levin classes, inter- 
sective Levin classes, as a source of semantic 
components associated with specific adjuncts. 
We also examine similar classes in Portuguese, 
and the predictive powers of alternations in this 
language with respect to the same semantic 
components. The difficulty of determining a 
suitable lexical representation becomes multi- 
plied when more than one language is involved 
and attempts are made to map between them. 
Preliminary investigations have indicated that 
a straightforward translation of Levin classes 
into other languages is not feasible (Jones et 
al., 1994), (Nomura et al., 1994), (Saint-Dizier, 
1996). However, we have found interesting par- 
allels in how Portuguese and English treat reg- 
ular sense extensions. 
2 Classifying verbs 
Two current approaches to English verb classi- 
fications are WordNet (Miller et al., 1990) and 
Levin classes (Levin, 1993). WordNet is an on- 
line lexical database of English that currently 
contains approximately 120,000 sets of noun, 
verb, adjective, and adverb synonyms, each rep- 
resenting a lexicalized concept. A synset (syn- 
293 
onym set) contains, besides all the word forms 
that can refer to a given concept, a definitional 
gloss and - in most cases - an example sentence. 
Words and synsets are interrelated by means 
of lexical and semantic-conceptual links, respec- 
tively. Antonymy or semantic opposition links 
individual words, while the super-/subordinate 
relation links entire synsets. WordNet was de- 
signed principally as a semantic network, and 
contains little syntactic information. 
Levin verb classes are based on the ability of 
a verb to occur or not occur in pairs of syntac- 
tic frames that are in some sense meaning pre- 
serving (diathesis alternations) (Levin, 1993). 
The distribution of syntactic frames in which a 
verb can appear determines its class member- 
ship. The fundamental assumption is that the 
syntactic frames are a direct reflection of the un- 
derlying semantics. Levin classes are supposed 
to provide specific sets of syntactic frames that 
are associated with the individual classes. 
The sets of syntactic frames associated with 
a particular Levin class are not intended to be 
arbitrary, and they are supposed to reflect un- 
derlying semantic components that constrain al- 
lowable arguments. For example, break verbs 
and cut verbs are similar in that they can all 
participate in the transitive and in the mid- 
dle construction, John broke the window, Glass 
breaks easily, John cut the bread, This loaf cuts 
easily. However, only break verbs can also occur 
in the simple intransitive, The window broke, 
*The bread cut. In addition, cut verbs can oc- 
cur in the conative, John valiantly cut/hacked 
at the frozen loaf, but his knife was too dull to 
make a dent in it, whereas break verbs cannot, 
*John broke at the window. The explanation 
given is that cut describes a series of actions di- 
rected at achieving the goal of separating some 
object into pieces. It is possible for these ac- 
tions to be performed without the end result 
being achieved, but where the cutting manner 
can still be recognized, i.e., John cut at the loaf. 
Where break is concerned, the only thing speci- 
fied is the resulting change of state where the 
object becomes separated into pieces. If the 
result is not achieved, there are no attempted 
breaking actions that can still be recognized. 
2.1 Ambiguities in Levin classes 
It is not clear how much WordNet synsets 
should be expected to overlap with Levin 
classes, and preliminary indications are that 
there is a wide discrepancy (Dorr and Jones, 
1996), (Jones and Onyshkevych, 1997), (Doff, 
1997). However, it would be useful for the 
WordNet senses to have access to the detailed 
syntactic information that the Levin classes 
contain, and it would be equally useful to have 
more guidance as to when membership in a 
Levin class does in fact indicate shared seman- 
tic components. Of course, some Levin classes, 
such as braid (bob, braid, brush, clip, coldcream, 
comb, condition, crimp, crop, curl, etc.) are 
clearly not intended to be synonymous, which 
at least partly explains the lack of overlap be- 
tween Levin and WordNet. 
The association of sets of syntactic frames 
with individual verbs in each class is not as 
straightforward as one might suppose. For in- 
stance, carry verbs are described as not taking 
the conative, *The mother carried at the baby, 
and yet many of the verbs in the carry class 
(push, pull, tug, shove, kick) are also listed in 
the push/pull class, which does take the cona- 
tive. This listing of a verb in more than one 
class (many verbs are in three or even four 
classes) is left open to interpretation in Levin. 
Does it indicate that more than one sense of 
the verb is involved, or is one sense primary, 
and the alternations for that class should take 
precedence over the alternations for the other 
classes in which the verb is listed? The grounds 
for deciding that a verb belongs in a particular 
class because of the alternations that it does not 
take are elusive at best. 
3 Intersective Levin classes 
We augmented the existing database of Levin 
semantic classes with a set of intersective 
classes, which were created by grouping to- 
gether subsets of existing classes with over- 
lapping members. All subsets were included 
which shared a minimum of three members. If 
only one or two verbs were shared between two 
classes, we assumed this might be due to ho- 
mophony, an idiosyncrasy involving individual 
verbs rather than a systematic relationship in- 
volving coherent sets of verbs. This filter al- 
lowed us to reject the potential intersective class 
that would have resulted from combining the re- 
move verbs with the scribble verbs, for example. 
The sole member of this intersection is the verb 
294 
draw. On the other hand, the scribble verbs 
do form an intersective class with the perfor- 
mance verbs, since paint and write are also in 
both classes, in addition to draw. The algorithm 
we used is given in Figure 1. 
1. Enumerate all sets S = {Cl,...,cn) of se- 
mantic classes such that Icl ¢3... f3 cnl _> e, 
where e is a relevance cut-off. 
2. For each such S = {cl,...,c~}, define an 
intersective class Is such that a verb v E 
Is iff v E cl f3 ... f3 cn, and there is no 
S' = {c~l,...,c~) such that S C S' and 
v e c~ f3... f3 c' m (subset criterion). 
Figure h Algorithm for identifying relevant 
semantic-class intersections 
We then reclassified the verbs in the database 
as follows. A verb was assigned membership in 
an intersective class if it was listed in each of 
the existing classes that were combined to form 
the new intersective class. Simultaneously, the 
verb was removed from the membership lists of 
those existing classes. 
3.1 Using intersective Levin classes to 
isolate semantic components 
Some of the large Levin classes comprise verbs 
that exhibit a wide range of possible semantic 
components, and could be divided into smaller 
subclasses. The split verbs (cut, draw, kick, 
knock, push, rip, roll, shove, slip, split, etc.) 
do not obviously form a homogeneous seman- 
tic class. Instead, in their use as split verbs, 
each verb manifests an extended sense that can 
be paraphrased as "separate by V-ing," where 
"V" is the basic meaning of that verb (Levin, 
1993). Many of the verbs (e.g., draw, pull, push, 
shove, tug, yank) that do not have an inherent 
semantic component of "separating" belong to 
this class because of the component of .force in 
their meaning. They are interpretable as verbs 
of splitting or separating only in particular syn- 
tactic frames ( I pulled the twig and the branch 
apart, I pulled the twig off (of) the branch, but 
not *I pulled the twig and the branch). The ad- 
junction of the apart adverb adds a change of 
state semantic component with respect to the 
object which is not present otherwise. These 
fringe split verbs appear in several other inter- 
sective classes that highlight the .force aspect of 
their meaning. Figure 2 depicts the intersection 
of split, carry and push/pull. 
"Split" Verbs 
Figure 2: Intersective class formed from Levin 
carry, push/pull and split verbs - verbs in 0 are 
not listed by Levin in all the intersecting classes 
but participate in all the alternations 
The intersection between the push/pull verbs 
of exerting force, the carry verbs and the split 
verbs illustrates how the force semantic compo- 
nent of a verb can also be used to extend its 
meaning so that one can infer a causation of 
accompanied motion. Depending on the par- 
ticular syntactic frame in which they appear, 
members of this intersective class (pull, push, 
shove, tug, kick, draw, yank) * can be used to 
exemplify any one (or more) of the the compo- 
nent Levin classes. 
1. Nora pushed the package to Pamela. 
(carry verb implies causation of accompa- 
nied motion, no separation) 
2. Nora pushed at/against the package. 
° Although kick is not listed as a verb of exerting force, 
it displays all the alternations that define this class. Sim- 
ilarly, draw and yank can be viewed as carry verbs al- 
though they are not listed as such. The list of members 
for each Levin verb class is not always complete, so to 
check if a particular verb belongs to a class it is better to 
check that the verb exhibits all the alternations that de- 
fine the class. Since intersective classes were built using 
membership lists rather than the set of defining alterna- 
tions, they were similarly incomplete. This is an obvious 
shortcoming of the current implementation of intersec- 
tive classes, and might affect the choice of 3 as a relevance 
cut-off in later implementations. 
295 
(verb of exerting force, no separation or 
causation of accompanied motion implied) 
3. Nora pushed the branches apart. 
(split verb implies separation, no causation 
of accompanied motion) 
4. Nora pushed the package. 
(verb of exerting force; no separation im- 
plied, but causation of accompanied motion 
possible) 
5. *Nora pushed at the package to Pamela. 
Although the Levin classes that make up an 
intersective class may have conflicting alterna- 
tions (e.g., verbs of exerting force can take the 
conative alternation, while carry verbs cannot), 
this does not invalidate the semantic regularity 
of the intersective class. As a verb of exerting 
force, push can appear in the conative alterna- 
tion, which emphasizes its force semantic com- 
ponent and ability to express an "attempted" 
action where any result that might be associ- 
ated with the verb (e.g., motion) is not nec- 
essarily achieved; as a carry verb (used with a 
goal or directional phrase), push cannot take the 
conative alternation, which would conflict with 
the core meaning of the carry verb class (i.e., 
causation of motion). The critical point is that, 
while the verb's meaning can be extended to 
either "attempted" action or directed motion, 
these two extensions cannot co-occur - they are 
mutually exclusive. However the simultaneous 
potential of mutually exclusive extensions is not 
a problem. It is exactly those verbs that are 
triple-listed in the split/push/carry intersective 
class (which have force exertion as a semantic 
component) that can take the conative. The 
carry verbs that are not in the intersective class 
(carry, drag, haul, heft, hoist, lug, tote, tow)are 
more "pure" examples of the carry class and 
always imply the achievement of causation of 
motion. Thus they cannot take the conative al- 
ternation. 
3.2 Comparisons to WordNet 
Even though the Levin verb classes are defined 
by their syntactic behavior, many reflect seman- 
tic distinctions made by WordNet, a classifica- 
tion hierarchy defined in terms of purely se- 
mantic word relations (synonyms, hypernyms, 
etc.). When examining in detail the intersec- 
tive classes just described, which emphasize not 
only the individual classes, but also their rela- 
tion to other classes, we see a rich semantic lat- 
tice much like WordNet. This is exemplified by 
the Levin cut verbs and the intersective class 
formed by the cut verbs and split verbs. The 
original intersective class (cut, hack, hew, saw) 
exhibits alternations of both parent classes, and 
has been augmented with chip, clip, slash, snip 
since these cut verbs also display the syntactic 
properties of split verbs. 
WordNet distinguishes two subclasses of cut, 
differentiated by the type of result: 
1. Manner of cutting that results in separa- 
tion into pieces (chip, clip, cut, hack, hew, 
saw, slash, snip), having cut, separate with 
an instrument as an immediate hypernym. 
2. Manner of cutting that doesn't separate 
completely (scrape, scratch), having cut 
into, incise as an immediate hypernym, 
which in turn has cut, separate with an in- 
strument as an immediate hypernym. 
This distinction appears in the second-order 
Levin classes as membership vs. nonmember- 
ship in the intersective class with split. Levin 
verb classes are based on an underlying lat- 
tice of partial semantic descriptions, which are 
manifested indirectly in diathesis alternations. 
Whereas high level semantic relations (syn- 
onym, hypernym) are represented directly in 
WordNet, they can sometimes be inferred from 
the intersection between Levin verb classes, as 
with the cut/split class. 
However, other intersective classes, such as 
the split/push/carry class, are no more con- 
sistent with WordNet than the original Levin 
classes. The most specific hypernym common 
to all the verbs in this intersective class is move, 
displace, which is also a hypernym for other 
carry verbs not in the intersection. In addition, 
only one verb (pull) has a WordNet sense cor- 
responding to the change of state - separation 
semantic component associated with the split 
class. The fact that the split sense for these 
verbs does not appear explicitly in WordNet 
is not surprising since it is only an extended 
sense of the verbs, and separation is inferred 
only when the verb occurs with an appropriate 
adjunct, such as apart. However, apart can also 
be used with other classes of verbs, including 
many verbs of motion. To explicitly list separa- 
296 
tion as a possible sense for all these verbs would 
be extravagant when this sense can be gener- 
ated from the combination of the adjunct with 
the force (potential cause of change of physical 
state) or motion (itself a special kind of change 
of state, i.e., of position) semantic component of 
the verb. WordNet does not currently provide 
a consistent treatment of regular sense exten- 
sion (some are listed as separate senses, others 
are not mentioned at all). It would be straight- 
forward to augment it with pointers indicating 
which senses are basic to a class of verbs and 
which can be generated automatically, and in- 
clude corresponding syntactic information. 
3.3 Sense extension for manner of 
motion 
Figure 3 shows intersective classes involving two 
classes of verbs of manner of motion (run and 
roll verbs) and a class of verbs of existence (me- 
ander verbs). Roll and run verbs have seman- 
tic components describing a manner of motion 
that typically, though not necessarily, involves 
change of location. In the absence of a goal or 
path adjunct they do not specify any direction 
of motion, and in some cases (e.g., float, bounce) 
require the adjunct to explicitly specify any dis- 
placement at all. The two classes differ in that 
roll verbs relate to manners of motion charac- 
teristic of inanimate entities, while run verbs 
describe manners in which animate entities can 
move. Some manner of motion verbs allow a 
transitive alternation in addition to the basic in- 
transitive. When a roll verb occurs in the tran- 
sitive (Bill moved the box across the room), the 
subject physically causes the object to move, 
whereas the subject of a transitive run verb 
merely induces the object to move (the coach 
ran the athlete around the track). Some verbs 
can be used to describe motion of both animate 
and inanimate objects, and thus appear in both 
roll and run verb classes. The slide class parti- 
tions this roll/run intersection into verbs that 
can take the transitive alternation and verbs 
that cannot (drift and glide cannot be causative, 
because they are not typically externally con- 
trollable). Verbs in the slide/roll/run intersec- 
tion are also allowed to appear in the dative 
alternation (Carla slid the book to Dale, Carla 
slid Dale the book), in which the sense of change 
of location is extended to change of possession. 
When used intransitively with a path prepo- 
sitional phrase, some of the manner of motion 
verbs can take on a sense of pseudo-motional 
existence, in which the subject does not actu- 
ally move, but has a shape that could describe 
a path for the verb (e.g., The stream twists 
through the valley). These verbs are listed in 
the intersective classes with meander verbs of 
existence. 
"Slide" Verbs 
R n Verbs 
? 
"Meander Verbs" 
Figure 3: Intersections between roll and run 
verbs of motion and meander verbs of existence 
4 Cross-linguistic verb classes 
The Portuguese verbs we examined behaved 
much more similarly to their English counter- 
parts than we expected. Many of the verbs 
participate in alternations that are direct trans- 
lations of the English alternations. However, 
there are some interesting differences in which 
sense extensions are allowed. 
4.1 Similar sense extensions 
We have made a preliminary study of the Por- 
tuguese translation of the carry verb class. As in 
English, these verbs seem to take different alter- 
nations, and the ability of each to participate in 
an alternation is related to its semantic content. 
Table i shows how these Portuguese verbs natu- 
rally cluster into two different subclasses, based 
on their ability to take the conative and apart 
alternations as well as path prepositions. These 
subclasses correspond very well to the English 
subclasses created by the intersective class. 
The conative alternation in Portuguese is 
mainly contra (against), and the apart alterna- 
tion is mainly separando (separating). For ex- 
ample, Eu puxei o ramo e o galho separando-os 
297 
English Portuguese Conat. Apart Path 
carry levar no no yes 
drag arrsStar no yes yes 
haul fretar no no yes 
heft levsntar corn 
dificuldade no no yes 
hoist icar no no yes 
lug levsr corn 
dificuldsde no no yes 
tote levar 
facilrnente no no yes 
tow rebocar no no yes 
shove ernpurrar corn 
violencia yes yes yes 
push ernpurrar yes yes yes 
draw puxar yes yes yes 
pull puxar yes yes yes 
kick chutar yes yes yes 
tug puxar corn 
forca yes yes yes 
yank arrancar yes yes yes 
Table h Portuguese carry verbs with their al- 
ternations 
(I pulled the twig and the branch apart ), and Ele 
empurrou contra a parede (He pushed against 
the wal O. 
4.2 Changing class membership 
We also investigated the Portuguese translation 
of some intersective classes of motion verbs. We 
selected the slide/roll/run, meander/roll and 
roll/run intersective classes. Most verbs have 
more than one translation into Portuguese, so 
we chose the translation that best described the 
meaning or that had the same type of arguments 
as described in Levin's verb classes. 
The elements of the slide/roll/run class are 
rebater (bounce), flutuar (float), rolar (rol 0 and 
deslizar (slide). The resultative in Portuguese 
cannot be expressed in the same way as in En- 
glish. It takes a gerund plus a reflexive, as in A 
porta deslizou abrindo-se (The door slid opening 
itselj~. Transitivity is also not always preserved 
in the translations. For example, flutuar does 
not take a direct object, so some of the alterna- 
tions that are related to its transitive meaning 
are not present. For these verbs, we have the in- 
duced action alternation by using the light verb 
fazer (make) before the verb, as in Maria fez o 
barco flutuar (Mary floated the boat). 
As can be seen in Table 2 the alternations for 
the Portuguese translations of the verbs in this 
intersective class indicate that they share simi- 
lar properties with the English verbs, including 
the causative/inchoative. The exception to this, 
as just noted, is flutuar (float). The result of 
this is that flutuar should move out of the slide 
class, which puts it with derivar (drift) and pla- 
nar (glide) in the closely related roll/run class. 
As in English, derivar and planar are not exter- 
nally controllable actions and thus don't take 
the causative/inchoative alternation common to 
other verbs in the roll class. Planar doesn't take 
a direct object in Portuguese, and it shows the 
induced action alternation the same way as flu- 
tuar (by using the light verb \]azer). Derivar is 
usually said as "estar a deriva" ("to be adrift"), 
showing its non-controllable action more explic- 
itly. 
5 Discussion 
We have presented a refinement of Levin classes, 
intersective classes, and discussed the potential 
for mapping them to WordNet senses. Whereas 
each WordNet synset is hierarchicalized accord- 
ing to only one aspect (e.g., Result, in the case of 
cut), Levin recognizes that verbs in a class may 
share many different semantic features, without 
designating one as primary. Intersective Levin 
sets partition these classes according to more co- 
herent subsets of features (force, force+motion, 
force+separation), in effect highlighting a lattice 
of semantic features that determine the sense of 
a verb. Given the incompleteness of the list of 
members of Levin classes, each verb must be 
examined to see whether it exhibits all the al- 
ternations of a class. This might be approxi- 
mated by automatically extracting the syntac- 
tic frames in which the verb occurs in corpus 
data, rather than manual analysis of each verb, 
as was done in this study. 
We have also examined a mapping between 
the English verbs that we have discussed and 
their Portuguese translations, which have sev- 
eral of the same properties as the corresponding 
verbs in English. Most of these verbs take the 
same alternations as in English and, by virtue 
of these alternations, achieve the same regular 
sense extensions. 
There are still many questions that require 
further investigation. First, since our experi- 
ment was based on a translation from English 
to Portuguese, we can expect that other verbs in 
Portuguese would share the same alternations, 
so the classes in Portuguese should by no means 
be considered complete. We will be using re- 
sources such as dictionaries and on-line corpora 
to investigate potential additional members of 
our classes. Second, since the translation map- 
pings may often be many-to-many, the alterna- 
298 
rebater flutuar 
(bounce) (float) 
dative yes 
*conative no 
caus./inch, yes 
middle yes 
accept, corer, yes 
caus./inch, yes 
resultative yes yes 
adject, part. :~es yes 
ind. action yes yes 
locat, invers, yes yes 
measure yes ye8 
*adj. perf. no no 
*cogn. object no no 
zero nom. yes yes 
Table 2: Portuguese slide/roll~run and 
tions may depend on which translation is cho- 
sen, potentially giving us different clusters, but 
it is uncertain to what extent this is a factor, 
and it also requires further investigation. In 
this experiment, we have tried to choose the 
Portuguese verb that is most closely related to 
the description of the English verb in the Levin 
class. 
We expect these cross-linguistic features to be 
useful for capturing translation generalizations 
between languages as discussed in the litera- 
ture (Palmer and Rosenzweig, 1996), (Copes- 
take and Sanfilippo, 1993), (Dorr, 1997). In 
pursuing this goal, we are currently implement- 
ing features for motion verbs in the English 
Tree-Adjoining Grammar, TAG (Bleam et al., 
1998). TAGs have also been applied to Por- 
tuguese in previous work, resulting in a small 
Portuguese grammar (Kipper, 1994). We in- 
tend to extend this grammar, building a more 
robust TAG grammar for Portuguese, that will 
allow us to build an English/Portuguese trans- 
fer lexicon using these features. 

References 
Tonia Bleam, Martha Palmer, and Vijay Shanker. 
1998. Motion verbs and semantic features in tag. 
In TAG-l--98, Philadelphia, PA. Submitted. 
Ann Copestake and Antonio Sanfilippo. 1993. Mul- 
tilingual lexical representation. In Proceedings of 
the AAAI Spring Symposium: Building Lexicons 
for Machine Translation, Stanford, California. 
Bonnie J. Dorr and Doug Jones. 1996. Acquisition 
of semantic lexicons: Using word sense disam- 
biguation to improve precision. In Proceedings of 
SIGLEX, Santa Cruz, California. 
Bonnie J. Dorr. 1997. Large-scale dictionary con- 
struction for foreign language tutoring and in- 
terlingual machine translation. Machine Trans- 
lation, 12:1-55. 
Doug Jones and Boyan Onyshkevych. 1997. Com- 
parisons of levin and wordnet. Presentation in 
working session of Semantic Tagging Workshop, 
ANLP-97. 
Douglas Jones, Robert Berwick, Franklin Cho, 
Zeeshan Khan, Karen Kohl, Naouky Nomura, 
Anand Radhakrishnan Ulrich Sanerland, and 
Bryan Ulicny. 1994. Verb classes and alternations 
in bangla, german, english, and korean. Technical 
report, Massachussets Institute of Technology. 
Karin Kipper. 1994. Uma investigacao de utilizacao 
do formalismo das gramaticas de adjuncao de ar- 
vores para a lingua portuguesa. Master's Thesis, 
CPGCC, UFRGS. 
B. Levin. 1993. English Verb Classes and Alterna- 
tions. 
J.B. Lowe, C.F. Baker, and C.J. Fillmore. 1997. A 
frame-semantic approach to semantic annotation. 
In Proceedings 1997 Siglex Workshop/ANLP97, 
Washington, D.C. 
G. Miller, R. Beckwith, C. Fellbaum, D. Gross, and 
K. Miller. 1990. Five papers on wordnet. Tech- 
nical Report 43, Cognitive Science Laboratory, 
Princeton University, July. 
Naoyuki Nomura, Douglas A. Jones, and Robert C. 
Berwick. 1994. An architecture for a universal 
lexicon: A case study on shared syntactic infor- 
mation in japanese, hindi, ben gall, greek, and en- 
glish. In Proceedings of COLING, pages 243-249, 
Santa Cruz, California. 
Martha Palmer and Joseph Rosenzweig. 1996. Cap- 
turing motion verb generalizations with syn- 
chronous tags. In AMTA-96, Montreal, Quebec, 
October. 
James Pustejovsky. 1991. The generative lexicon. 
Computational Linguistics, 17(4). 
Patrick Salnt-Dizier. 1996. Semantic verb classes 
based on 'alternations' and on wordnet-like se- 
mantic criteria: A powerful convergence. In Pro- 
ceedings the Workshop on Predicative Forms in 
Natural Language and Lexieal Knowledge Bases, 
Toulouse, France. 
