Using Subcategorization to Resolve Verb Class Ambiguity 
Maria Lapata 
School of Cognitive Science 
Division of Informatics 
University of Edinburgh 
2 Buccleuch Place 
Edinburgh EH8 9LW, UK 
mlap@cogsci.ed.ac.uk 
Chris Brew 
HCRC Language Technology Group 
Division of Informatics 
University of Edinburgh 
2 Buccleuch Place 
Edinburgh EH8 9LW, UK 
chrisbr@cogsci.ed.ac.uk 
Abstract 
Levin's (1993) taxonomy of verbs and their classes 
is a widely used resource for lexical semantics. In 
her framework, some verbs, such as give exhibit no 
class ambiguity. But other verbs, such as write, can 
inhabit more than one class. In some of these am- 
biguous cases the appropriate class for a particular 
token of a verb is immediately obvious from inspec- 
tion of the surrounding context. In others it is not, 
and an application which wants to recover this infor- 
mation will be forced to rely on some more or less 
elaborate process of inference. We present a simple 
statistical model of verb class ambiguity and show 
how it can be used to carry out such inference. 
1 Introduction 
The relation between the syntactic realization of a 
verb's arguments and its meaning has been exten- 
sively studied in Levin (1993). Levin's work re- 
lies on the hypothesis that "the behavior of a verb, 
particularly with respect to the expression and in- 
terpretation of its arguments, is to a large extent 
determined by its meaning" (Levin, 1993, p. 1). 
Verbs which display the same diathesis alterna- 
tions-alternations in the realization of their argu- 
ment structure-are assumed to share certain mean- 
ing components and are organized into a semanti- 
cally coherent class. 
As an example consider sentences (1)-(3) 
taken from Levin. Example (1) illustrates the 
causative/inchoative alternation. Verbs undergoing 
this alternation can be manifested either as transi- 
tive with a causative reading (cf. (la)) or as intransi- 
tive with an inchoative reading (cf. (lb)). Examples 
(2) and (3) illustrate the dative and benefactive al- 
ternations respectively. Verbs which license the for- 
mer alternate between the prepositional frame NP- 
V-NP-PPto (cf. (2a)) and the double object frame 
V-NP-NP (cf. (2b)), whereas verbs which undergo 
the latter alternate between the double object frame 
(cf. (3a)) and the prepositional frame NP-V-NP- 
PPJ~r (cf. (3b)). 
(1) a. Janet broke the cup. 
b. The cup broke. 
(2) a. Bill sold a car to Tom. 
b. Bill sold Tom a car. 
(3) a. Martha carved the baby a toy. 
b. Martha carved a toy for the baby. 
Verbs like crack and chip pattern with break in li- 
censing the causative/inchoative alternation and are 
associated with the semantic class of BREAK verbs. 
Verbs make and build behave similar to carve in 
licensing the benefactive alternation and are mem- 
bers of the class of BUILD verbs, whereas sell and 
give undergo the dative alternation and participate 
in the GIVE class. By grouping together verbs which 
pattern together with respect to diathesis alterna- 
tions Levin defines approximately 200 verb classes, 
which she argues reflect important semantic regu- 
larities. 
2 Motivation 
Levin provides an index of 3,024 verbs for which 
she lists the semantic classes and diathesis alterna- 
tions. The mapping between verbs and classes is 
not one-to-one. Of the 3,024 verbs which she cov- 
ers, 784 are listed as having more than one class. 
Even though Levin's monosemous verbs outnumber 
her polysemous verbs by a factor of nearly four to 
one, the total frequency of the former (4,252,715) 
is comparable to the total frequency of the latter 
(3,986,014). This means that close to half of the 
cases processed by a hypothetical semantic tagger 
would manifest some degree of ambiguity. The fre- 
quencies are detailed in table 1 and were compiled 
from a lemmatized version of the British National 
Corpus (BNC), a widely distributed 100 million 
word collection of samples of written and spoken 
English (Burnard, 1995). 
266 
Classes \[ Verbs I BNCfrequency 
1 2,239 4,252,715 
2 536 2,325,982 
3 173 738,854 
4 43 395,212 
5 23 222,747 
6 7 272,669 
7 2 26,123 
10 1 4,427 
Table 1: Polysemous verbs according to Levin 
I O0 
90 
80 
70 
6o 
~ so 
30 
20 
I0 
0 1 2 ,3 4 5 6 7 8 9 10 11 12 
Number of alternations 
Classes 
1 
2 
3 
4 
5 
6 
7 
Figure 1: Relation between number of classes and 
alternations 
Furthermore, as shown in figure 1, the number 
of alternations licensed by a given verb increases 
with the number of classes it inhabits. Consider 
for example verbs participating in one alternation 
only: of these, 90.4% have one semantic class, 8.6% 
have two classes, 0.7% have three classes and 0.3% 
have four classes. In contrast, of the verbs licensing 
six different alternations, 14% have one class, 17% 
have two classes, 12.4% have three classes, 53.6% 
have four classes, 2% have six classes and 1% has 
seven classes. 
Palmer (1999) and Dang et al. (1998) argue that 
the use of syntactic frames and verb classes can sim- 
plify the definition of different verb senses. Beyond 
this, we claim that information about the argument 
structure of a polysemous verb can often help dis- 
ambiguating it. 
Consider for instance the verb serve which is a 
member of four classes: GIVE, FIT, MASQUERADE 
and FULFILLING. Each of these classes can in turn 
license four distinct syntactic frames. As shown in 
the examples I below, in (4a) serve appears ditran- 
sitively and belongs to the semantic class of GIVE 
verbs, in (4b) it occurs transitively and is a mem- 
ber of the class of FIT verbs, in (4c) it takes the 
predicative complement as minister of the interior 
and is a member of MASQUERADE verbs. Finally, 
in sentence (4d) serve is a FULFILLING verb and 
takes two complements, a noun phrase (an appren- 
ticeship) and a prepositional phrase headed by to. 
In the case of verbs like serve we can guess their 
semantic class solely on the basis of the frame with 
which they appear. 
(4) a. I'm desperately trying to find a venue for 
the reception which can serve our guests 
an authentic Italian meal. 
b. The airline serves 164 destinations in over 
75 countries. 
c. Jean-Antoine Chaptal was a brilliant 
chemist and technocrat who served 
Napoleon as minister of the interior from 
1800 to 1805. 
d. Before her brief exposure to pop stardom, 
she served an apprenticeship to a still-life 
photographer. 
But sometimes we do not have the syntactic infor- 
mation that would provide cues for semantic disarn- 
biguation. Consider sentence (5a). The verb write 
is a member of three Levin classes, two of which 
(MESSAGE TRANSFER, PERFORMANCE) take the 
ditransitive flame NP-V-NP-NP. In this case we 
have the choice between the "message transfer" 
reading (cf. (5a)) and the "performance" reading 
(cf. (Sb)). This is an instance of the common prob- 
lem of inferring the value of a hidden variable (in 
this case the "true class" of a particular instance 
of write). The same situation arises with the verb 
phone which is listed as a GET verb and an INSTRU- 
MENT OF COMMUNICATION verb and in both cases 
can take the frame NP-V-NP-NP. In sentence (5c) 
the preferred reading is that of "get" instead of "in- 
strument of communication" (cf. sentence (5d)). 
(5) a. A solicitor wrote him a letter at the air- 
port. 
b. I want you to write me a screenplay called 
"The Trip". 
1 Unless stated otherwise the example sentences were taken 
from the BNC and simplified for clarification purposes. 
267 
c. I'll phone you a taxi. 
d. As I entered the room I wished I'd thought 
of phoning a desperate SOS to James. 
The objective of this paper is to address the 
verb class disambiguation problem by developing 
a probabilistic framework which combines linguis- 
tic knowledge (i.e., Levin's classification) and frame 
frequencies acquired from the BNC. Our initial ex- 
periments focus on the syntactic frames characteris- 
tic for the dative and benefactive alternations (cf. ex- 
amples (2) and (3)). These frames are licensed by 
a fairly large number of classes: 19 classes license 
the double object frame, 22 license the NP-V-NP- 
PPto frame and 14 classes license the NP-V-NP PPfi~r 
frame. The semantic and syntactic properties 
of these alternations have been extensively studied 
and are well understood (see Levin (1993) and the 
references therein). Furthermore, they are fairly pro- 
ductive and one would expect them to be well rep- 
resented in a large corpus. 
In section 3 we describe the statistical model and 
the estimation of the various model parameters, sec- 
tion 4 presents some preliminary results and sec- 
tion 5 contains some discussion and concluding re- 
marks. 
3 The Model 
We view the choice of a class for a polysemous 
verb in a given frame as the joint probability 
P(verb,frame, class) which we rewrite using the 
chain rule in (6). 
(6) P(verb,frame, class) = P(verb) 
e (frame lverb) P (class I verb, frame) 
We also make the following independence assump- 
tion: 
(7) P(classlverb,frame) ~ P(class\[frame) 
The independence assumption reflects Levin's hy- 
pothesis that the argument structure of a given verb 
is a direct reflection of its meaning. Accordingly we 
assume that the semantic class determines the argu- 
ment structure of its members without making ref- 
erence to the individual verbs. By applying Bayes 
Law we write P(classlframe) as: 
(8) P(class\[frame)= P (frame lclass) P (class) P (frame ) 
By substituting (7) and (8) into (6), 
P(verb, class,frame) can be written as: 
(9) P(verb,frame, class) 
P (verb) P (frame l verb) P (frame lclass) P (class) 
P ~rame ) 
We estimate the probabilities P(verb), 
P(framelverb), P(framelclass) and P(class) 
as follows: 
(10) P(verb) f(verb) f (verbi) 
i 
(11) P(framelverb) 
(12) P(framelclass) 
(13) P(class) .~ 
f (verb,frame) 
f (verb, framei) 
i 
f (class,frame) 
f (class, framei) 
i 
f (class) 
Y~. f (classi) 
i 
(14) P(frame) ,~ f (frame) f {framei) 
i 
It is easy to obtain f(verb) from the lemmatized 
BNC. For the experiments reported here, syntactic 
frames for the dative and benefactive alternations 
were automatically extracted from the BNC using 
Gsearch (Keller et al., 1999), a tool which facilitates 
search of arbitrary POS-tagged corpora for shallow 
syntactic patterns based on a user-specified context- 
free grammar and a syntactic query. The acquisition 
and filtering process is detailed in Lapata (1999). 
We rely on Gsearch to provide moderately accu- 
rate information about verb frames in the same way 
that Hindle and Rooth (1993) relied on Fidditch to 
provide moderately accurate information about syn- 
tactic structure, and Ratnaparkhi (1998) relied on 
simple heuristics defined over part-of-speech tags 
to deliver information nearly as useful as that pro- 
vided by Fidditch. We estimated f(verb,frame) as 
the number of times a verb co-occurred with a par- 
ticular frame in the corpus. 
We cannot read off P(frame\[class) from the cor- 
pus, because it is not annotated with verb classes. 
Nevertheless we can use the information listed in 
Levin with respect to the syntactic frames exhib- 
ited by the verbs of a given class. For each class 
268 
Class Frames 
MAN NER NP-V-NP-PP#om, NP-V-NP, 
NP-V-PPat, NP-V-NP-PRED 
ACCOMPANY NP-V-NP, NP-V-NP-PPt,, 
THROW NP-V-NP-NP, NP-V-NP-PPtoc, 
NP-V-NP-PP#om-PPto, NP-V-NP, 
NP-V-NP-PPto, NP-V-NP-PPar, 
PERFORMANCE NP-V, NP-V-NP, NP-V-NP-NP, 
NP-V-NP-PPto, NP-V-NP-PP¢~r, 
NP-V-NP 
G I v E NP-V-NP-PPto, NP-V-NP-NP 
CONTRIBUTE NP-V-NP-PPto 
Table 2: Sample of verb classes and their syntactic 
frames 
we recorded the syntactic frames it licenses (cf. ta- 
ble 2). Levin's description of the argument struc- 
ture of various verbs goes beyond the simple list- 
ing of their subcategofization. Useful information 
is provided about the thematic roles of verbal argu- 
ments and their interpretation. Consider the exam- 
ples in (15): in (15a) the verb present is a member of 
the FULFILLING class and its theme is expressed by 
the prepositional phrase with an award, in (15b) the 
PP headed by with receives a locative interpretation 
and the verb load inhabits the SPRAY/LOAD class, 
whereas in (15c) the prepositional phrase is instru- 
mental and hit inhabits the HIT class. None of the 
information concerning thematic roles was retained. 
All three classes (FULFILLING, SPRAY/LOAD and 
HIT) were assigned the frame NP-V-NP-PPwith'. 
(15) a. 
b. 
C. 
John presented the student with an award. 
John loaded the truck with bricks. 
John hit the wall with a hammer. 
Because we didn't have corpus counts for the 
quantity f(class,frame) we simply assumed that 
all frames for a given class are equally likely. 
This means, for instance, that the estimate for 
P(NP-V-NP-NPtolGIvE) is ½ and similarly the es- 
timate for P(NP-VIPERFORMANCE ) is ~ (cf. ta- 
ble 2). This is clearly a simplification, since one 
would expect f(class,frame) to be different for dif- 
ferent corpora, and to vary with respect to class size 
and the frequency of class members. 
In order to estimate P(class) we first estimate 
f(class) which we rewrite as follows: 
(16) f (class) = E f (verbi, class) 
i 
Class size(class) p(class\[amb_class) f (verb, class) 
THROW 27 0.40 7783.6 
SEND 20 0.27 5253.9 
GIVE 15 0.20 3891.8 
MARRY 10 0.13 2529.6 
Table 3: Estimation of f(verb, class) for the verb 
pass 
The estimate of f(verb, class) for monosemous 
verbs reduces to the count of the verb in the cor- 
pus. Once again we cannot estimate f(verb, class) 
for polysemous verbs directly. All we have is the 
overall frequency of a given verb in the BNC and 
the number of classes it is a member of according to 
Levin. We rewrite f(verb, class) as: 
(17) f (verb, class) = f (verb)p(classlverb) 
We approximate p(classlverb) by collapsing across 
all verbs that have the appropriate pattern of ambi- 
guity: 
(18) f (verb, class) ~ f (verb)p(classlamb_class) 
Here amb_class, the ambiguity class of a verb, is 
the set of classes that it might inhabit. 2 We collapse 
verbs into ambiguity classes in order to reduce the 
number of parameters which must be estimated: we 
certainly lose information, but the approximation 
makes it easier to get reliable estimates from limited 
data. In future work we plan to use the EM algo- 
rithm (Dempster et al., 1977) to uncover the hidden 
class, but for the present study, we simply approxi- 
mate p(classlamb_class) using a heuristic based on 
class size: 
size(class) (19) p(classlamb_class) 
size(c) 
c ~ amb~'lass 
For each class we recorded the number of its mem- 
bers after discarding verbs whose frequency was 
less than 1 per 1M in the BNC. This gave us a first 
approximation of the size of each class. We then 
computed, for each polysemous verb, the total size 
of the classes of which it was a member. We calcu- 
lated p(classlamb_class) by dividing the former by 
the latter (cf. equation (19)). We obtained an esti- 
mate for the class frequency f(class) by multiply- 
ing p(classlamb_class) by the observed frequency 
of the verb in the BNC (cf. equation (18)). 
2Our use of ambiguity classes is inspired by a similar use in 
HMM based part-of-speech tagging (Kupiec, 1992). 
269 
~e.+05 - 
4e+05 - 
u~ 
2e.H)5 - 
® 
® 
® 
I !: ..... ~: :;N 
o N N 
"¢ E " 
Figure 2: The ten most frequent classes 
120 
1oo 
~ so 
~ 60 
E 
~ 40 
77- 
20 ~ ,;. 
I 
t 
Figure 3: 
0. 
Z I 
g 
Z 
m 
Ten most frequent frames in Levin 
As an example consider the verb pass which has 
the classes THROW, SEND, GIVE and MARRY. The 
respective p(classlamb_class) for these classes are 
27 20 15 and l0 By multiplying these by the fre- 72' 72' 72 ~" 
quency of pass in the BNC (19,559) we obtain 
the estimates for f(verb, class) given in table 3. 
Note that simply relying on class size, without re- 
gard to frequency, would give quite different results. 
For example the class of MANNER OF SPEAKING 
verbs has 76 members, of which 30 have frequen- 
cies which are less than 1 per 1M, and is the sev- 
enth largest class in Levin's classification. Accord- 
ing to our estimation scheme MANNER OF SPEAK- 
ING verbs are the 116th largest class. The estimates 
for the ten most frequent classes are shown in fig- 
ure 2. 
The estimation process described above in- 
volves at least one gross simplification, since 
p(classlamb_class) is calculated without reference 
to the identity of the verb in question. For any 
two verbs which fall into the same set of classes 
p(classlamb_class) will be the same, even though 
one or both may be atypical in its distribution across 
the classes. Furthermore, the estimation tends to 
favour large classes, again irrespectively of the iden- 
tity of the verb in question. For example the verb 
carry has three classes, CARRY, FIT and COST. In- 
tuitively speaking, the CARRY class is the most fre- 
quent (e.g., Smoking can impair the blood which 
carries oxygen to the brain, I carry sugar lumps 
around with me). However, since the FIT class 
(e.g., Thameslink presently carr/es 20,000 passen- 
gers daily) is larger than the CARRY class, it will be 
given a higher probability (0.45 versus 0.4). This is 
clearly wrong, but it is an empirical question how 
much it matters. 
Finally, we wanted to estimate the probability 
of a given frame, P(frame). We could have done 
this by acquiring Levin compatible subcategoriza- 
tion frames from the BNC. Techniques for the auto- 
matic acquisition of subcategofization dictionaries 
have been developed by Manning (1993), Bfiscoe 
and Carroll (1997) and Carroll and Rooth (1998). 
But the present study was less ambitious, and nar- 
rowly focused on the frames representing the da- 
tive and the benefactive alternation. In default of the 
more ambitious study, which we plan for the future, 
the estimation of P(frame) was carried out on types 
and not on tokens. The mapping of Levin's linguis- 
tic specifications into surface syntactic information 
resulted in 79 different frame types. By counting the 
number of times a given frame is licensed by several 
semantic classes we get a distribution of frames, a 
sample of which is shown in figure 3. 
The probabilities P(frmnelclass) and 
P(framelverb) will be unreliable when the 
frequency estimates for f(verb,frame) and 
f(class,frame) are small, and ill-defined when 
the frequency estimates are zero. Following 
Hindle and Rooth (1993) we smooth the ob- 
served frequencies in the following way, where 
f(V,frame) = ~i.f(verbi,frame), f(V) = 
270 
~i f (verbi), f (C,frame) : ~i f (classi,ft~me) 
and f(C) = ~i f(classi). We redefine the 
probability estimates as follows: 
(20) P (framel verb) 
f(V,Jmme) f (verb,frame) + f(v) 
f (verb,framei) + 1 
i 
(21 ) P (framelclass) 
f(C,/~ame) f (class,frame) + f(c) 
f (class,framei) + 1 
i 
When f(verb,frame) is zero, the estimate used 
is proportional to the average f(V.frame) f(v) across all 
verbs. Similarly, when f(class,frame) is zero, our 
estimate is proportional to the average f(c.l'~ame) f(C) 
across all classes. We don't claim that this scheme is 
perfect, but any deficiencies it may have are almost 
certainly masked by the effects of approximations 
and simplifications elsewhere in the system. 
4 Results 
We evaluated the performance of the model on all 
verbs listed in Levin which are polysemous and take 
frames characteristic for the dative and benefactive 
alternations. This resulted in 154 verbs which take 
the NP-V-NP-NP frame, 135 verbs which take the 
NP-V-NP-PPw frame and 84 verbs which take the 
NP-V-NP-PPj~,r frame. The verbs were all polyse- 
mous and had an average of 3.8 classes. Each class 
had an average of 3.4 frames. Furthermore, we di- 
vided these verbs in two categories: verbs which can 
be disambiguated solely on the basis of their frame 
(e.g., serve; category A) and verbs which are gen- 
uinely ambiguous, i.e., they inhabit a single frame 
and yet can be members of more than one semantic 
class (e.g., write; category B). 
The task was the following: given that we know 
the frame of a given verb can we predict its se- 
mantic class? In other words by varying the class 
in the term P(verb,frame, class) we are trying to 
see whether the class which maximizes it is the one 
predicted by the lexical semantics and the argument 
structure of the verb in question. 
For the verbs belonging to category A (306 in 
total) we used Levin's own classification in eval- 
uation. The model's performance was considered 
correct if it agreed with Levin in assigning a verb 
the appropriate class given a particular frame. For 
class ambiguous verbs (category B) we compared 
the model's predictions against manually annotated 
data. Given the restriction that these verbs are se- 
mantically ambiguous in a specific syntactic frame 
we could not simply sample from the entire BNC, 
since this would decrease the chances of finding the 
verb in the frame we are interested in. Instead, for 
31 class ambiguous verbs we randomly selected ap- 
proximately 100 tokens from the data used for the 
acquisition of frame frequencies for the dative and 
benefactive alternation. Verbs with frame frequency 
less than 100 were not used in the evaluation. 
The selected tokens were annotated with class in- 
formation by two judges. The judges were given an- 
notation guidelines but no prior training. We mea- 
sured the judges' agreement on the annotation task 
using the Kappa coefficient (Siegel and Castellan, 
1988) which is the ratio of the proportion of times, 
P(A), that k raters agree to the proportion of times, 
P(E), that we would expect the raters to agree by 
chance (cf. (22)). If there is a complete agreement 
among the raters, then K = 1, whereas if there is no 
agreement among the raters (other than the agree- 
ment which would be expected to occur by chance), 
then K = 0. 
P(A) - P(E) (22) K- 
1 - P(E) 
We counted the performance of our model as cor- 
rect if it agreed with the "most preferred", i.e., most 
frequent verb class as determined in the manually 
annotated corpus sample by taking the average of 
the responses of both judges. 
We also compared the results for both categories 
to a naive baseline which relies only on class in- 
formation and does not take subcategorization into 
account. For a given polysemous verb, the baseline 
was computed by defaulting to its most frequent 
class, where class frequency was determined by the 
estimation procedure described in the previous sec- 
tion. 
As shown in table 4, in all cases our model out- 
performs the baseline. It achieves a combined pre- 
cision of 91.8% for category A verbs. One might 
expect a precision of 100% since these verbs can 
be disambiguated solely on the basis of the frame. 
However, the performance of our model is less, 
mainly because of the way we estimated the terms 
P(class) and P(frame\[class): we overemphasize 
the importance of frequent classes without taking 
into account how individual verbs distribute across 
classes. 
The model achieves a combined precision of 
83.9% for category B verbs (cf. table 4). Further- 
271 
II Category A 
Frame \[Verbs Baseline 
NP-V-NP-NP 123 61.8% 
NP-V-NP-PPto 113 67.2% 
NP'V-NP-PPfor 70 70% 
combined 
Category B 
Model Verbs Baseline Model 
87.8% 14 42.8% 85.7% 
92% 15 73.4% 86.6% 
98.5% 2 0% 50% 
II 306 165.7% 191.8% \[31 161.3% 183.9% 
Table 4: Model accuracy against baseline 
Verb Frame Preferences 
save NP-V-NP-NP 
call NP-V-NP-NP 
write NP-V-NP-NP 
make NP-V-NP-NP 
extend NP-V-NP-PPto 
present NP-V-NP-PPto 
take NP-V-NP-PPj~,~ 
produce NP-V-NP-PPfbr 
GET, BILL 
GET, DUB 
MESSAGE TRANSFER, PERFORMANCE 
DUB, BUILD 
FUTURE HAVING, CONTRIBUTE 
FULFILLING, REFLEXIVE APPEARANCE 
STEAL, PERFORMANCE 
PERFORMANCE, CREATE 
Table 5: Random sample of eight verbs and their semantic preferences as ranked by the model 
more, our model makes interesting predictions with 
respect to the semantic preferences of a given verb. 
In table 5 we show the class preferences the model 
came up with for eight randomly selected verbs 
(class preferences are ranked from left to right, with 
the leftmost class being the most preferred one). Ta- 
ble 6 summarizes the average class frequencies for 
the same eight verbs as assigned to corpus tokens 
by the two judges together with inter-judge agree- 
ment (K). The category OTHER is reserved for cor- 
pus tokens which either have the wrong frame or 
for which the classes in question are not applicable. 
In general agreement on the class annotation task 
was good with Kappa values ranging from 0.68 to 
1. As shown in table 6, with the exceptions of call 
and produce the model's predictions are borne out 
in corpus data. 
5 Discussion 
Verb 
save 
call 
GET 
64 
GET 
2 
Class 
BILL 
25 
DUB 
94 
write M. TRANS. PERF. 
54 19 
make DUB BUILD 
59 20 
extend FUT. HAV. CONTR. 
50 37 
present FULFIL. R. APP. 
79 18 
take PERF. CREATE 
52 13 
produce PERF. CREATE 
8 91 1 
I K 
OTHER 0.74 
11 
OTHER 0.82 
4 
OTHER 0.85 
18 
OTHER 0.78 
21 
OTHER 0.71 
13 
OTHER 0.94 
3 
OTHER 0.77 
33 
OTHER i 0.73i 
Table 6: Random sample of eight verbs and their 
semantic preferences as ranked by two judges 
This paper explores the degree to which syntactic 
frame information can be used to disambiguate verb 
semantic classes. In doing so, we cast the task of 
verb class disambiguation in a probabilistic frame- 
work which exploits Levin's semantic classification 
and frame frequencies acquired from the BNC. The 
approach is promising in that it achieves high preci- 
sion witha simple model and can be easily extended 
to incorporate other sources of information which 
can influence the class selection process (i.e., selec- 
tional restrictions). 
The semantic preferences which we generate can 
be thought of as default semantic knowledge, to be 
used in the absence of any explicit contextual or 
lexico-semantic information to the contrary (cf. ta- 
ble 5). Consider the verb write for example. The 
272 
model comes up with an intuitively reasonable rank- 
ing: we more often write things to people ("message 
transfer" reading) than for them ("performance" 
reading). However, faced with a sentence like Max 
wrote Elisabeth a book pragmatic knowledge forces 
us to prefer the "performance" reading versus the 
the "message transfer" reading. In other cases the 
model comes up with a counterintuitive ranking. For 
the verb call, for instance, the "get" reading (e.g., I 
will call you a cab) is preferred over the more natu- 
ral "dub" reading (e.g., John called me a fool). 
We still rely heavily on the verb class informa- 
tion provided by Levin. But part of original aim 
was to infer class information for verbs not listed 
by Levin. For such a verb, P(class), and hence 
P(verb,frame, class) will be zero, which is not 
what we want. Recent work in computational lin- 
guistics (e.g., Schfitze (1993)) and cognitive psy- 
chology (e.g., Landauer and Dumais (1997)) has 
shown that large corpora implicitly contain seman- 
tic information, which can be extracted and manipu- 
lated in the form of co-occurrence vectors. The idea 
would be to compute the centroid (geometric mean) 
of the vectors of all members of a semantic class. 
Given an unknown verb (i.e., a verb not listed in 
Levin) we can decide its semantic class by compar- 
ing its semantic vector to the centroids of all seman- 
tic classes. We could (for example) determine class 
membership on the basis of the closest distance to 
the centroid representing a semantic class (cf. Patel 
et al. (1998) for a proposal similar in spirit). Once 
we have chosen a class for an unknown verb, we are 
entitled to assume that it will share the broad syn- 
tactic and semantic properties of that class. 
We also intend to experiment with a full scale 
subcategorization dictionary acquired from the 
BNC. We believe this will address issues such as: 
(a) relations between frames and classes (what are 
the frames for which the semantic class is predicted 
most accurately) and (b) relations between verbs 
and classes (what are the verbs for which the seman- 
tic class is predicted most accurately). We also plan 
to experiment with different classification schemes 
for verb semantics such as WordNet (Miller et al., 
1990) and intersective Levin classes (Dang et al., 
1998). 

References 

Ted Briscoe and John Carroll. 1997. Automatic 
extraction of subcategofization from corpora. In 
Proceedings of the 5th ACL Conference on Ap- 
plied Natural Language Processing, pages 356- 
363, Washinton, DC. 

Lou Bumard, 1995. Users Guide for the British Na- 
tional Corpus. British National Corpus Consor- 
tium, Oxford University Computing Service. 

Glenn Carroll and Mats Rooth. 1998. Valence 
induction with a head-lexicalized PCFG. In 
Proceedings of the 3rd Conference on Empir- 
ical Methods in Natural Language Processing, 
Granada. 

Hoa Trang Dang, Karin Kipper, Martha Palmer, and 
Joseph Rosenzweig. 1998. Investigating regu- 
lar sense extensions based on intersective Levin 
classes. In Proceedings of the 17th International 
Conference on Computational Linguistics and 
36th Annual Meeting of the Association for Com- 
putational Linguistics, pages 293-299, Montrral. 

A. E Dempster, N. M. Laird, and D. B. Rubin. 
1977. Maximum likelihood for incomplete data 
via the EM algorithm. Journal of the Royal Sta- 
tistical Socie~, 39(2):1-38. 

Donald Hindle and Mats Rooth. 1993. Structural 
ambiguity and lexical relations. Computational 
Linguistics, 19(1): 103-120. 

Frank Keller, Martin Corley, Steffan Corley, 
Matthew W. Crocker, and Shaft Trewin. 1999. 
Gsearch: A tool for syntactic investigation of 
unparsed corpora. In Proceedings of the EACL 
Workshop on Linguistically Interpreted Corpora, 
Bergen. 

Julian Kupiec. 1992. Robust oart-of-speech tag- 
ging using a hidden Markov model. Computer 
Speech and Language, 6(3):225-242. 

Thomas K. Landauer and Susan T. Dumais. 1997. 
A solution to Plato's problem: The latent seman- 
tic analysis theory of acquisition, induction and 
representation of knowledge. Psychological Re- 
view, 104(2):211-240. 

Mafia Lapata. 1999. Acquiring lexical generaliza- 
tions from corpora: A case study for diathesis 
alternations. In Proceedings of the 37th Annual 
Meeting of the Association for Computational 
Linguistics, College Park, MA. 

Beth Levin. 1993. English Verb Classes and Alter- 
nations: A Preliminary. Investigation. University 
of Chicago Press, Chicago. 

Christopher D. Manning. 1993. Automatic acquisi- 
tion of a large subcategorization dictionary from 
corpora. In Proceedings of the 31st Annual Meet- 
ing of the Association for Computational Linguis- 
tics, pages 235-242, Columbus, OH. 

George A. Miller, Richard Beckwith, Christiane 
Fellbaum, Derek Gross, and Katherine J. Miller. 
1990. Introduction to WordNet: An on-line lexi- 
cal database. International Journal of Lexicogra- 
phy, 3(4):235-244. 

Martha Palmer. 1999. Consistent criteria for sense 
distinctions. Computers and the Humanities, to 
appear. 

Malti Patel, John A. Bullinaria, and Joseph E Levy. 
1998. Extracting semantic representations from 
large text corpora. In John A. Bullinaria, D. W. 
Glasspool, and G. Houghton, editors, In Proceed- 
ings of the 4th Workshop on Neural Computa- 
tion and Psychology, pages 199-212. Springer, 
Berlin. 

Adwait Ratnaparkhi. 1998. Unsupervised statisti- 
cal models for prepositional phrase attachment. 
In Proceedings of the 7th International Confer- 
ence on Computational Linguistics, pages 1079-1085. 

Hinrich Schtitze. 1993. Word space. In 
Stephen. Jos6 Hanson, Jack D. Cowan, and 
C. Lee Giles, editors, Advances in Neural In- 
formation Processing Systems, pages 895-902. 
Morgan'Kaufmann, San Mateo, CA. 

Sidney Siegel and N. John Castellan. 1988. Non 
Parametric Statistics for the Behavioral Sciences. 
McGraw.-Hill, New York. 
