Lexi(:al Knowledge Acquisition ti'om Bilingual Col'por& 
Takehito UTSURO* Yuji MATSUMOTO Makoto NAGAO 
l)cpt, of Electrical Engineering, Kyoto (iniversity 
Yoshida-honmachi, Sakyo-Ku, Kyoto, 606, Japan 
utsuro(@kuee.kyoto-u.ac.j p 
Abstract 
I)br practical research in natnral language processing, 
it is indisl)ensM)le to develop a large scale semantic 
dictionary for computers. It is cspeciany important 
to improve thc tcclmiqucs tbr compiling semantic dic- 
tionaries ti'orn natural language texts such as those in 
existing human dictionaries or in large corpora, llow- 
ever, there are at least two ditlicultics in analyzing 
existing texts: tbe l)roblem of syntactic ambiguities 
and the probtcm of polysemy. Our approaclL to solve 
these difficulties is to make use of translation exam- 
pies in two distinct languages that have (lnite different 
syntactic structures and word meanings. The roe.son 
we took this at)preach is that in many cases both syn: 
tactic aLrd semantic ambignitics arc resolved by com- 
paring analyzed resnlts from botb languages. In this 
paper, we propose a method Ibr resolving the syntac- 
tic ambiguities of translation cxaml>lcs of bilingual 
corpora and a method for acquiring lexical knowl- 
edge, such as ease frames of verbs and attribute sets 
el noons. 
1 Introduction 
It has become widely accel)ted that developing a large 
scale semantic dictionary is indispensable to future 
natural language research. ILL recent years, several 
research activities for compiling selnantic dictionar- 
ies tot natural language processing have been uudcr- 
taken One of the approaches in this research is at- 
tempts to compile dictionaries by band. Japan Elec- 
tronic Dictionary Research Institute (El)R.) is now 
compiling conceptual dictionaries\[5\] by hand with 
the help of software tools. \[nformation-4echnology 
Promotion Agency (IPA), Japan, has also compiled 
IPA Lexicon of the Japanese Language for computers 
(II'AL)\[4\]. IPAL has 861 entries for basic Jalranese 
verbs. Cyc project attempts to assend)le a mas- 
sive knowledge base covering human common-sense 
knowledge\[7\]. IIowever, this approach sailors from 
*The authol~ would like to t}mak the editorial staff of Ko- 
dazm|m for permission tO use the data of Jalmnese-12)nglidt dic- 
tionaa'y, arm also thank l)r. Shouichi YOKOYAMA, I,',TL, and 
Prof. l\[ozumi TANAKA and Dr. '\['akenobu TOKUNA(;A, 
Tokyo hmtitute of Teclmology, for providing us the data of 
Jal)ane~e-l~nglish dictionary. This work is partly supported by 
the Grants from Ministry of Education, #032,15103. 
probh'.Ins socb as a huge alnount of manila\[ labor, 
difficulties in extending tile dictionaries, unstable re- 
milts, and so forth. 
Anothcr approach is to compile dictionaries us- 
ing some teclxmques of lexical knowledge acquisition. 
One ~nch approach is to extract hierarclfical rela- 
tions or it thesanrtm of conceptual items froln hunLall 
dictionaries in an automatic way. q)surrnnaru et el. 
studied to construct a t}LeSaLLrlIs of nominal concepts 
from noun detinitions\[t3\], qbmiara et al. also ex- 
tracted snperordinatc-subordmatc relation between 
verbs from the defining sentences in IPAL\[12\]. lie 
sidcs these rcseasches, there are other several research 
activitics tbr lexical knowledge acquisition, which syn- 
tactically anMyze the sentences m large corpora and 
attcmpt to extract lcxical knowledge from statisti- 
cal data \[3\] \[1\]. Most of the works undertake shallow 
analysis of texts and they extract only superticial lex- 
ical information. 
For the development of tile techniques of knowledge 
acquisition from natural language texts, it is very im- 
portant to improve the httter approach of cornpiling 
semantic dictionaries by comimter l)rograuL~. Ilow- 
ever, there are at least two basic difficulties in this 
at)preach 
1. Tire i~robh~m (ff syntactic ambiguities 
When analyzing a sentence., syntactic ambiguities 
often remain. So i~ is not easy to obtain correct 
parsed results automatically. 
2. The, probh~rrr of polyue,my 
it often happens that one word has several mean- 
ings and corre.sponrls to ,~cveral concepts. So it is 
not easy to associate one sm'fa~e word with olle 
correct conceptHal item. 
Our approach to solve these diiliculties is to make 
use of translatitm cxarnples in two distinct languages 
that have quite different syntactic structures and 
word mf~anings (such as English and Japanese), and 
to c(nnt~are analyzed results from each language, h| 
many (:asc~, the two languagcs }Lave different types 
of syntactic ambiguities, anti comparison of syntactic 
structures of both bmguagcs helps to resolve the am- 
biguities. Also, a pair of bilingually equivalent snrface 
words helps to a~'4ociate tile words with conceptual 
Ac~s DECOLING-92, NANTIiS, 23-28 AOIJT 1992 5 8 1 l'~oc. OF COL\]NG-92. NANTES. AUG. 23-28, 1992 
words helps to associate the words with conceptual 
items, because the intersection of conceptual items 
that each surface word has could be considered as 
one conceptual item\[ll\] \[2\]. \["or example, in tire case 
of the translation example given in Example 1, both 
syntactic and semantic ambiguities are resolved. 
Example 1 
E: I hung my coat on the hook. 
J: ~:L (I) ;~ (topic) ~2~ (coat) ~ (ca.se-m~trker) 
~'5" (hook) lZ (case-marker) zi'$~'f: (hung)o 
1. Syntactic disumbiguation 
The English sentence in Example 1 is syntacti- 
cally ambiguous because the prepositional phrase 
"on the hook" can modify both the verb "hung" 
aad tim noun phrase "my coat" using grammat- 
ical knowledge only. On the other band, in the 
Japanese sentence, the phrase "7)~ ~', Is_" can mod- 
ify nothing but the verb "~t;tf:". Thus, if 
knowledge about word equivalence pairs such as 
( I, ~, ), (hung, ~'~t t: ), (coat, _t-.;a ), (hook, ~' g ) 
are available from bilingual dictionaries, the ambi- 
guity of pp-attachment is resolved by syntactically 
matching the structures of the two sentences. 
2. Semantic disambiguation 
The verb "~)~l~ ~" in tile Japanese sentence is a 
typical Japanese polyserny. This verb has six sub- 
entries in a Japanese dictionary that has about 
70,000 entries, and ten English equivalent verbs 
( "hang", "spend", "play", etc.) in a Japanese- 
English dictionary that has about 50,000 entries. 
So, it is not easy to associate the surface word 
"~qJ'~" with its exact meaning. Ilowever, with 
the translation examl)le , the corresponding En- 
glish verb such ms "hang" helps to find the mean- 
irrg of the Japanese verb "7~19 ~,5''. 
In this paper, we propose a method for resolving 
the syntactic ambiguities of translation examples in 
bilingual corpora and a method for acquiring lexi- 
cal knowledge, such as case frames of verbs and at- 
tribute sets of nouns. In our framework, first a pair 
of sentences of both languages are syntactically ana- 
lyzedtand translated into feature descriptions, which 
represent dependency structures of the pbrases in the 
sentences. Although feature descriptions are gener- 
ated by grarnmatical knowledge only, they are quite 
suitable to represent case frames of verbs. Then these 
feature descriptions of the two languages are com- 
pared, or unified, using knowledge about word equiv- 
alence from bilingual dictionaries. In this matching 
process, one word in the English sentence could be 
eqnivalent to several words in the translated Japanese 
1Tbe Japanese morphological analyT~r lm.s 14 part of apeech 
and about 36,000 words. The Englisb dictionary contains 
about 55,DO0 words. The current Japanese and English grana- 
mar~ consist of 85 DCG rules aald 135 DCG rul~-s. 
sentence. Also one word in the Japanese sentence 
could be equivalent to several words in the translated 
English sentence. In order to realize the matching 
process between two languages including these sev- 
eral word equivalence cases, we introduce a unifica- 
tion algorithm based on sets of compatible pairs of 
atomic values and feature labels in Chapter 2. 
In Chapter 3, we statistically evaluated the process 
of syntactic disambiguation. The success ratio of dis- 
ambiguation is about 63~68 % for translation exam- 
pies in a Japanese-English dictionary. At present, we 
have already collected about 50,000 translation exam- 
pies from a machine readable Japanese-English dic- 
tionary (Kodansha Japanese-English Dictionary \[10\]) 
and an English learners' textbook. We have extracted 
case frames for several verbs as a simple experiment. 
The results are described in Chapter 4. 
2 Unification of Feature De- 
scriptions of Two Languages 
2.1 Unification based on Sets of Com- 
patible Pairs of Features and Val- 
ues 
In our framework of sentence analysis, a sentence in 
each language is parsed and translated into feature 
descriptions, which represent dependency structures 
of the phrases in the sentence. Ill this section, we ba- 
sically use and extend Kasper and Rounds' notation 
of feature description logic (FDL \[6\]) to describe our 
unification algorithm of feature descriptions, except 
that we don't use path equivalence. 
When unifying feature descriptions of two lan- 
guages, knowledge about word equivalence taken 
from bilingual dictionaries is used to decide whether 
all atomic value of one language is compatible with 
an atomic value of the other language. This is also 
the casc with feature labels. Knowledge about word 
equivalence from bilingual dictionaries can be re- 
garded as knowledge about compatibility of atomic 
values and feature labels of feature descriptions. 
From this standpoint, we introdnce a unification al- 
goritlHn based on sets of compatible pairs of atomic 
values and feature labels. 
Data Structure 
Let A and L be sets of symbols used to denote atomic 
values and feature labels. Let CA and CL be sets of 
compatible pairs of atomic values and feature labels. 
That is, (/A is the set of pairs of atomic values such as 
(ai,aj)(al, aj ~_ A), where al and u i are consistent and 
mfifiable, and Ct. is the set of pairs of feature labels 
like {li,lj)(li,lj C L), where li and lj are consistent 
AC't'ES DE COLING-92, NANTES, 23-28 AOt~rr 1992 5 8 2 PROC. OF COLING-92, NAN'rES, AUG. 23-28, 1992 
and unitlable 2'3. 
The syntax for formulas of the FDL with Sets of 
Compatible Pairs (FDLC) is given below. 
NIL denoting no information 
TOP denoting inconsistent information 
a where a E A, to describe atolnic values 
(ai, aj) where ai, aj E A and (ctl, aj) E CA, 
to describe pairs of atomic values 
1 : ¢ where I E L and ¢ E FI)I,C, 
to describe structures in which the feature 
labeled by / ha.s a value described by ¢ 
(li,lj) : ¢ whereli,b (5 L and (l,,Ij) C- CL 
attd ¢ (: FDLC, 
to describe structures in which the feature 
labeled by (li, Ij) hmu a value described by ¢ 
¢ A ¢ where ¢, ~b G FDLC 
Unification Algorithm 
Because of the compatibility scts, there is not nec- 
essarily a unique most general unifier of two feature 
descriptions. When applying this algorithm to unify 
fe.aturc descriptions between two languages, we col 
lect all possible unified feature descriptions and lind 
the most overlapping Ulfifier by a scoring function, 
which is introdneed later. The following detinition of 
UNIFY returns one possible unified feature descrip- 
tion. We collect all possible nnitied feature descrip- 
tions. 
Function UNIFY(f ,g) returns one possible 
unified feature description: 
where f attd g are featur)e descmptions. 
1. If f =NIL, then return g 
2. Else if g = N1L, then return f 
3. Else if f = TOP or g = TOP, 
then return "1'01" 
4. Else iff, gEAtJCA and f--9 
then return f(: g) 
5. Else if f,g E A, 
if (f,g) G CA, tt ...... t ..... (f,g) 
else return TOP 
end. 
6. Else if f = 1 : a I attd g = l : u s, 
and IE LUG'L, 
if( alg := )UNIFY(a:,a~), 
then return I : al9 
else return "FOP 
end. 
~These compatibility sets do not necessarily define equiv 
alence relations of atomic vtdu~ and feature labels, i.e., ttley 
do not satisfy the trmmitive ~ld symmetric laws. They race 
rellexive, and (a,a) a~td (l,l) are identified ~s a and 1. 
a In fact, in the case of tile tulificatlon of feature descriptions 
of two languages, ai of (ai, aj)(~ CA) is an atomic value of ol~e 
language and a) is aa atomic value of the other lmlguage. This 
is also the case with I i gild 13 of (It, 1~)(~ CI. ), 
7. Else iff=l!:a! and g=l u:at, 
and (11, lg) (~ CL 
and ( aI~ := )UNIFY(ay,a~), 
then return (I),, lg) : aI~ 
8. Elseiff=flAf2 
and ( ..~ h, f,., g~ ~t, := )UNIFY-CONJ(f,g) 
and ( h .... )UNIFY(f,,g,), 
then return h A h~ 
9. Else if g = 9a A g2, then return UNIFY(g, f) 
1{), Else return f A g 
eltd. 
Function UNIFY-CONJ(f,g) retnrns one 
possihle 34uple of feature descriptions << 
h, fr, gr ~-': where f and g are feature descrip- 
tions, and h is a unified feature description, 
and fr,gr are r~t parts of f,g that are not 
used to generate h. 
1. if f -- f, A f~, 
( .~ h, f,, g,. ) :=)UNIFY-CONJ(f~, g) 
and return ~ h, f,. A f~, g~ Y~ 
or 
( ~ h,f,-,9, ~ :=)UNIFY-CONJ(f~,.q) 
and return ~ h, fl A fr,g,- Y~" 
2. Elscifg=glAg2 
and ( 42 h,g~,f~ ?~t,:-:)UNII"Y-CONJ(g,f) 
then return ,~ h, f~, 9," 
3. Else ( h :~ )UNWY(f, 9) 
and return ,( h, NIL, NIL ~t, 
cud, 
2.2 Unification of Feature Descrip- 
tions of Two Languages 
Feature Descriptions of translation examples of both 
languages are generated by syntactic analysis. A 
translation example is given in Example 2. 
Example 2 
E: I wrote it letter with a pencil. 
J: ~l, (1) t~t (topic) ~'~ (pencil) "if" (case-marker) 
:/:~i; (letter) ~ (caae-marker) ~'l= (wrote)o 
From the English sentence of this example, two fea- 
ture descriptions below are generated because of the 
ambiguity caused by pp-attachment. 
pred : write 
tertsc : past 
,.bj: \[ ,,~e,,: x \] 
\[ p.,l :,e.. 1 
' L spee : ,~ j 
w.h: \[ prig: Ve,,e" \] 
L spec :. j 
pred : write 
tense : past 
s.bj: \[ prea: l \] 
obj : spee : a 
with : pred : pencil 
spec : tt 
AcrEs DE COL1NG-92, NANTES, 23-28 hOGT 1992 5 8 3 PROC. Ol: COLING-92, NA~CrES. AUG. 23-28, t992 
From the Japanese sentence, the following single fea- 
ture description is generated. 
tense : past 
t:~ : pred : ?eL \] 
pred : ~ \] 
~¢ pred : ~ \] 
Set of Compatible Pairs of Atomic Values 
Knowledge about word equivalence is extracted from 
bilingual dictionaries m order to construct CA. First, 
for each word in the English sentence, equivalent 
Japanese words are extracted from English-Japanese 
dictionaries, and for each word in the Japanese sen- 
tence, equivalent English words are extracte.d from 
Japanese-English dictionaries 4, Using this knowl- 
edge, any possible pairs of equivalent cotttent words s 
that are included in the original sentences are col- 
lected, and CAD, the set of these equivalent (i.e. 
coml)atible ) word pairs, is constructed. Then for all 
other content words WND~, s in the English ~ntenee 
and WN1)Jap in the Japanese sentence, any possible 
pairs (WN:)¢~g, WNDiap) are collected, which com- 
prise CAN9. Finally, CA is defined ms CA:) U CAND. 
In the case of Example 2, CA:), CAN:) and CA are 
shown below. CA~9 and CAND are constructed only 
for the content words, so ill this ease CaN9 is ~ (an 
empty set). 
CAn = {(write, ~ (), (I, ~1,), (letter, :~\]~ ), (pencil, ~t~)}, 
CAN\]) = ~, CA = CAD tJ CAND 
Set of Compatible Pairs of Feature Labels 
In our framework of unification between two lan- 
guages, we assmne that the set of compatible pairs 
of feature labels, CL, is constructed based on sta- 
tistical data. That is, each feature label pair (li,lj) 
in CL has a probability plj(O < Pij <_ 1) calculated 
from statistical data. This Pij represents the proba- 
bility that the semantic role of feature Ii in a specific 
feature description of one lamguage is the same as 
that of feature l.i ill another specific feature descrip- 
tion of the other language. For exaurple, for a specific 
English Japanese verb pair (write, ~- ~ ), the feature 
label pair (sub j, ¢)¢ ) is ,~ssumed to have a probabil- 
ity P,ubL ~" And for anotlmr English--Japanese verb 
pair (read, ~t2 ), ttle feature label pair (subj, :b ¢ ) is 
assmned to have another probahility qsubj, h'. 
Since we are at the starting point of our project 
of lexical knowledge acquisition, we initially assign 1 
to tire probability of each feature label pair, except 
4At pre~ent, we use a Japan~e-English dictionary only, which has about 50,0(}0 entries. 
5Words are divided into two categories: content words mid 
fmlctional word~. Content words are ones which can be the 
head of a phrase, such ms i1o1111$ and verbs. 
for pairs that are known not to have ttle same case 
role from some grammatical knowledge. These ex- 
ceptional pairs are not contained ill CL, i.e., tlmir 
probabilities are 0. In fact, for the purpose of lexical 
knowledge acquisition, it is sufficient to assume the 
probability as 1 or 0, because we need credible results 
for extracting lexical knowledge about the usages of 
words. 
The Most Overlapplng Unifier 
The scoring function SCOR.E(h) calculates the va- 
lidity of a unified feature description h. This func- 
tion returns a 2-tnple of real numbers s, (xl,x2) 
(xl,x2 E R(set of real numbers)), where xl is the 
number of word pairs extracted from bilingual dictio- 
naries and contained ill the unified feature descrip- 
tion, on the other hand x~ is tile number of word 
pairs aLso contained in the unified feature descrip- 
tion but not extracted from bilingual dictionaries. 
More precisely, xl corresponds to tile number of word 
pairs (Wo,~9 , WDjop) in the unified feature descrip- 
tion that are elements of CAD, and x~ corresponds 
to the number of word pairs (WND~,s, WNOj,p) in 
the unified feature description that are elements of 
CAN D . 
The order among scores is defined as follows: 
{xt,x2) is greater tban (Yl,~) 
iff. xl >yl or (xt =yt,x2 >y2) 
The most overlapping unifiers are the ones with the 
greatest score. The complete definition of the scoring 
function is given below. 
Function SCORE(h) returns (xl, x~) (xl, x2 (5 
R(set of real numbers)): 
where h is a unified feature description. 
1. If h E CAJg, then return (1, 0) 
2. Else if h E CAND, then return (0, 1) 
3. Else ifh=l:a whereICLuCz and 
a E A U Ca and SCORE(a) = (x,,x2), 
then return 
(scortEL(1) × ~l,SCOltE~(t) x ~) 
4. Else if h = hi A h~ where hi, h2 E FDI,C 
and SCORE(h~)= (2:11 , ZI2 ) 
and SCORE(h~)= ( ....... ), 
then return (xll + x2~, zl2 -t- x22) 
5. Else return (0,0) 
end. 
Function SCOREL(I) returns the probability 
of l: wherelc LUCL 
1. If I E L, then return 1 
2. If/E CL, then return the probability of I 
eSince the probability of a feature label pair is l or 0, Xl 
and x 2 ate integers at pre~ellt. 
ACTrS n~: COLING-92. NANTES, 23-28 AO~r 1992 5 8 4 PROC. ol: COLING-92, NANTEs, AUG. 23-28. 1992 
Example 
The results of unification and scoring of Example 2 
are as below. 
score = (4, 0) 
pred : (write, ~ ( ) 
tense : past 
(o~j, \[ spec : a \] 
score = (3, 0) 
pred : (write, ~ < ) 
tense : past (~,,bj, ~ ): \[ 
pr,a: (x,~,)\] 
| spec : a J 
<o~j, ~ ): \[ ,..h \[ ,,~.1 : pe,,.l \] 
L " L spec : a j -¢: \[ 
prea : ~,i \] 
Tt~e prepositional phrase "with a pencil" modifies 
the verb "wrote" m the upper feature description. 
The score of tile upper feature description is greater 
than that of tile lower one. So in this ease, the upper 
one is regarded as tile correct ease frame example for 
tile pair (write, ~" < ). 
3 Syntactic Disambiguation: 
Experiment and Evaluation 
in order to evaluate how well syntactic ambiguities of 
translation examples are resolved, we made all exper- 
iment of syntactic disambiguation using 189 transla- 
tion examples extracted from a J apanese-English dic- 
tionary. Firstly, each sentence of a translation exam- 
pie is syntactically analyzed and translated into fea- 
ture descriptions. For 44 translation examples, syn- 
tactic analysis of tile Japanese or English sentence 
is faile.d. For those which are successfully analyzed, 
the average number of feature descriptions generated 
from one scntcncc is 4.4 for Japanese and 17.1 for En- 
glish. Secondly, these feature descriptions are unified. 
After this process of syntactic disamhiguation, from 
86 translation examples, a uniquc ee~sc framc of the 
unified verb pair of Japanese and English is acquired. 
Calculating from this result, the success ratio of ac- 
quiring unified case frames of verbs, (the number of 
translation examples such that a unique unified case 
frame of verbs is acquired from each translation exam- 
pie)/ (tile uumher of translation examples such that 
each sentence is successfully analyzed), is 86/145 = 
59.3%. And the success ratio of syntactic disambigua- 
tion, (tile number of sentences such that a unique 
ease frame of the verb is acquircd from more than 
one feature descriptions)/ (tile number of sentences 
such that more than one feature descriptions are orig- 
inally generated), is 70/103 = 68.0% for Japanese, 
and 84/133 = 63.2% for English. 
4 Lexical Knowledge Acquisi- 
tion of Verbs 
4.1 Acquiring Case Frames of Verbs 
As described ill 2.2, a feature description unified be- 
tween English aud Japanese is as below. 
pred : (write, ~ < ) 
tense : past 
obj ¢~ pred : (letter ~5.\]~ ) 
F pred <p,,,ea, ) I 
( h -c ) / .wtt , " : L spec : a J 
This feature description tells that tile verbal con- 
cept represented hy tile pair of the English verb 
"tv~te" and the Japane~qe verb "~ <" have at least 
three eases that are marked by some syntactic in- 
formation mid some surface functional words such 
as (subj, *2 ), (obj, ~ ), (with, T' ). it also tells that 
each case takes a certain nominal coueept represented 
by tile pair of English and Japanese words, such as 
U, *h >, <fetter, ~;:~ ), (pe,leit, ~ ). Once a large 
amount of this kind of data is collected, statistical 
data ahout case frames of verbs eaal he extracted, 
making use of a thesaurus of nominM concepts 7. In 
the remainder of this section, we will illustrate a gen- 
eral procedure for acquiring case frames of verbs. 
Lct us start with a collection of a large amount 
of unified feature descriptions like above for a specific 
Japanesc verb V~. Suppose that we want to get possi- 
ble case frames of this verb. By a case frame, we mean 
something tikc a feature description for this verb, con- 
sisting of surface cases each of which is marked hy a 
postpositional particlc p~ and some specific semantic 
categories taken from a thesaurus like BGI\[. Usually, 
a verh has several distinct case frames. However, it is 
not easy to extract those case frames automatically 
only from the collected unified feature descriptions. 
So the system finds critical points to distinguish pos- 
sible case frames for a verh using some heuristics, 
then it asks tile human instructor whether the dis- 
tinctions of ease frames arc correct. These heuristics 
and human interactions arc smmnarized as follows. 
7At present, ~m oiL-line thesaurus called 'Bunrui Goi 
Hyou'(BGH)\[8\] is available for Japanese. BGII has a six- 
layered abstraction hierarchy mrd more t|mat 60,OOO words are 
assigned at the leaves. At the presettt stage, it is ntot cer- 
tain whether this the~sautim is reliable enouglt for our initial 
research target of acquiring case frames of verbs. It is, how- 
ever~ the most precise and broad coveri|kg 3apsmeae thesaurus 
obtahtable for us, currently. 
ACRES DE COLING-92, NANTES, 23-28 ho~r 1992 5 8 5 PRO(. OF COLING-92, NAN'I'ES, AUG. 23-28, 1992 
Heuristics 
1. Semantic Categor'y in a Thesaurus 
First, collect the nouns marked by pj in a fea- 
ture description of the verb Vj from the set of 
unified feature descriptions. Then mark each col- 
lected noun in the thesaurus. If the most specific 
common layer of the marked nouns is low enough, 
then we assume that the case marked by pj takes 
a noun of the semantic category that corresponds 
to that layer. But if the most specific common 
layer is higher than a predetermined layer s, the 
information provided by that layer is too general 
for tile semantic categories of the case marked by 
pj. For instance, it is quite rare that both an ani- 
mate concept and an abstract concept can be the 
subject of a certain verb. Such a case strongly 
suggests that the verb has at least two distinct 
conceptual meanings or two distinct case frames. 
It then becomes necessary to classify the marked 
nouns in the thesaurus. 
2. Bilingual Intersection of Concepts 
Some of the heuristics come from the advan- 
tages of bilingual intersection of concepts, which 
we have already shows in Chapter 1 as seman- 
tic disambiguation. For a Japanese verb Vj and 
its case marked by a postpositional particle p j, 
suppose that unified feature descriptions such 
as \[ pred:(VEl,Vj) , (IEI,pj):{NE1,NJI) \] and 
\[ pred:(VE2 , Vj), (IE2,pJ):(NE2,NJ2) \] are oh- 
tained. Both of these two feature descriptions 
have a feature label pj for Vj. llowever, if VE1 
and V~2 are different verbs or IEl and IE2 are 
different feature labels, these two feature descrip- 
tions may be classified into different case frames 
of the verb Vj. 
3. Correlation of Cases 
Another heuristics are related to sentence pat- 
terns of verbs. Sometimes the ease marked by 
pj has a correlation with other eases in sentence 
patterns. If the correlations between cases are de- 
tected, then it helps tile classification, and some 
sentence patterns (or c0.se frames) of the verb Vj 
will be aeqnired. 
Human Interactions 
As described above, the system can find critical points 
to distinguish possible case frames for a verb by those 
heuristics. The system, however, cannot determine 
the distinction only with positive data collected from 
examples. The main purpose of human interaction 
is to obtain negative examples. The system asks the 
human instructor whether a case marked by p J1 and 
another case marked by P J2 call co-occur or not. If 
STile \])redetermined layers depend on tile thesaurus we are 
dealing with. 
Table 1: Semantic Marker of IPAL 
CON concrete ABS abstract 
ANI animal ACT action 
HUM human MEN mental 
ORG organization LIN linguistic products 
PLA plant CHA character 
PAR parts REL relation 
NAT natural LOC location 
PRO products TIM time 
QUA quantity 
I PHE I pt .......... II DIV I dl ..... 
Table 2: Acquired Case Slots for '~ < (write)" 
Case Slots Sere. Mark. Freq. Examples 
(subj, ~;;t - ~,¢ ) HUM 95 $/~ (l) 
(obj, t:t • ¢ ) REL, ~ (letter), 
(\[subj, passive\], QUA, 153 ~fi~/(name) 
I$ • h~) L1N 
(with, "~ ) PRO 10 "~ 5" (per,) 
(in, ~" ) L1N, ~.~ (kanji) 
ItEL 28 Jt~ (form) 
(on, ~:- ) PRO 16 i~ (paper) 
!to, iS. ) 11UM 13. ~ (father). 
they cannot co=occur, then the system learns that Vj 
}lets at least two sentence patterns (or case frames) 
and that one of them has the case marked by P J1 and 
tile other has the case marked by P J2- An example 
of human interactions of this type is shown in next 
section. 
It is often said that hand-made semantic dictio- 
nary contains quite unstable data, which means that 
it strongly depends on the human composer. In or- 
der to acquire stable lexicat knowledge base, we de- 
cided to limit hmnan interactions to yes-no type of 
questions and answers, such that the system asks the 
human instructor whether something is true or false 
so that he can answer only yes or no. 
4.2 Examples and Evaluations 
Wc have collected about 50,000 translation exam- 
pies from a machine readable dal)anese-English dic- 
tionary and an English learners' textbook. In this 
bilingual corpus, about 70 distinct Japanese verbs ap- 
pear in more than 100 examples. We have obtained 
unified feature descriptions for several verbs which 
appeared more than 200 times. From them we have 
gotten some case frames. In this experiment we used 
the set of semantic markers defined in IPAL \[4\], listed 
in Table 1. 
Table 2 shows the case slots of "~ < (write)" ex- 
tracted from 207 translation examples. In the process 
of extraction, bilingual feature label pairs are quite 
uscfut to find different case slots that are marked by 
the same postpositional particle in Japanese. In order 
to acquire ease fralne.s of tile verb "~ < (write)" from 
Ac'u~s DE COLING-92, NAtCrF.S, 23-28 Ao(rr 1992 5 8 6 PROC. OF COLING-92, NANTES, AUG. 23-28, 1992 
'Fable 3: Acquired Case Frames for "~-\[ < (wr;le)" 
(7~Lse Frame I Ca.~e Frlmm 2 
15 (on) PRO V- (to) HUM 
l~t . :6¢ (sub3) IIUM ~;t • fie (subj) HUM 
~) REL, l;~ " ~ (obj) 
J3.. ~ ( QUA, t:\]: • fit ( LIN 
\[subj,passive\]) LIN \[subj,p,,ssive\]) 
"~" (with) ~ (with) 
-e (i,,) -e (i,,) 
the extracted cm~e slots, ttle systenr ~sks the human 
instructor about the pcx~sibi\[ities of tile co-occurrence 
of the case slots that do not cc.occur in the trans 
lation examples by composing saml)le phr,'~ses. The 
questions and answers are as follows. 
QUESTION 1 : 
C'a, I say 
,,..'t Y (pen) -r. (with) l~.~,l~ (English) "C (in) ~ < {write)" ? 
....... YES. 
QUESTION 2 : 
Cat, I say 
"2 -- I-" U.a,'d) ~5 (o,,) 5t \[I,,the,') ~5 (t,,) ~ < (,mr, O'' 
..... NO. 
The postpositional particle "~:" is used to mark two 
different cases of the verb "~ < (write)" in Japm~ese 
sentences. One of them represents things on which 
smnething is written like in "wrile something on , 
sh~:et of paper", and the other reI)resents someone to 
whom a correspondence is written, like ill "wtalc a 
letter to a lover". The difference of these two usages 
is clear by tit(: bilingual feature label pairs (on, ~= ) 
and (to, {~ ). 'File human instructor answers that only 
these two ease slots cannot co-occur. Then two case 
frames are obtained as in Table 3. 
This simple experiment suggests that it is quite 
possible to acquire case frames of verbs from bilingual 
corpora if enough translation examples are available. 
Actually, on tim assumption that 200 translation ex- 
amples arc necessary for acquiring case framcs of onc 
verb, 100,000 translation examples are necessary for 
70 verbs. If a bilingual corpus of 1,000,000 transla- 
tion examples is obtained, it is possible to compile 
a semantic dictionary with the same scale as IPAI, 
through a little interaction with a human instructor 
for each verb. Wc think it possible to construct a 
bilingual corl)us of that scale or more in the near fit- 
lure, 
5 Concluding Remarks 
We haw~ proposed a method for resolving the syntac- 
tic ambiguities of translation examples of bilingual 
corpora and a method for acquiring case frames of 
verbs. At present, we are extending our prototype 
system for acquiring case frames of verbs, attd the 
detail of the extended system will be reported in the 
future. We believe that the I)roposed method is appli- 
cable to sew:ral otller problenrs as well. One of them 
is to acquire features of nominal concepts. We are at 
the moment looking at some specitie nominal expres- 
sion "A q) B" in Japanese, corresponding literally to 
"I1 of A" in English. That expression specifies a vari- 
ety of relationships of noun phrases, which are often 
stated in different expressions in English. They will 
help to acquire typical attributes of nominal concepts 
fl'om bilingual corpora. Our ntethod is also useful to 
collect parsed traamlation examples tbr example-based 
translation \[9\] attd to acquire translation patterns be- 
tween two languages. 

References 

\[l\] \[h'cnt, M.: "Automatic Acquisition of Subcatego- 
rization Frames from Untagged Text", f>roc, of the 
29th Anuual Meeti.9 of the AUL, 1991. 

\[2\] l)ag;~n, I., ltai, A. and Schwall, IJ.: "Two l,an- 
guages are More hlformative Than One", Proc. o\] 
the '29th Annual Meetin9 af the ACL, 1991. 

\[3\] ltindle, 1).: "Noun Classification from Predicate 
Argument Structures ~, Proc. o.\[ the 28th Annual 
Meeting of the ACL, 1990. 

\[4\] Information-tedmology Promotiml Agency, Japan: 
IPA Lexicon of the Japane,~e Language for comput- 
c,-s IPAL (Basic Vt:rbs),(i, Japanese), 1987. 

\[5\] Japan Electric l)ictionary Research Institute, Ltd.: 
Com:vptual Dictionary, 2rid. Edltion,(in J~paneae), 
TI/ 012, 1989. 

\[6\] Kasper, It. and ltouuds, W.: "A Logical Semantics 
for l"e~Lture Structures", Plot. o\] the ~th Anuual 
Meeting of the ACL, 19811. 

\[7\] Lenat, 1). et a/.: lluildin 9 Large Ktaowledye-based 
Systems, Addison-Wesley, 1990. 

\[8\] National Language Research Institute: Word List 
by Sema~dic Principles,tin Japanese), Syuei Syup- 
pan, 1964. 

\[9\] Sato, S. and Nagao, M.: "Memory-based Trans- 
lation", Proc. o~ the 12th International Conference 
o,~ Computational Linguistics, Ilelsinki, Finland, 
1990. 

\[10\] Shimizu, M. and Narita, N., cds.: Japanese-Enqlish 
Dictiotaary, Kodansha Gakujutsu t~,unko, 1979. 

\[11\] Tokunaga. T. and Tanaka, I1.: "Thc Aatmmttic 
Extraction of Conceptual Items from Bilingual Dic- 
tionaries', (in Japanese), Journal of Japan Society 
for Artificial bttelligence, Vol.6, No.2, 1991. 

\[12\] Tomiura, V., Ititaka, T. and Yoshida, S.: "Extract- 
ing the Superordinate Subordinate Relation be- 
tween Verbs from Definition Sentences in Jat)anese 
l)ietionaties", (in Japanese), Journal o\] In\]orma- 
lion Processing, 1PSJ, Vol.32, No.l, 1991. 

\[13\] Tsurumaru, H., qkkesita, K., ltami, K., Yanagawa, 
T. and Yoshida, S.: "An Approach to Thesaurus 
Construction",(in Japanese), H*SJ-WGNL, 83-16, 
1991. 
