A Method of Utilizing Domain and Language specific 
Constraints in Dialogue Translation 
Masami SUZUKI 
ATR Interpreting Telephony Research Laboratories 
2-2 Hikaridai, Seika-cho, Soraku-gun, Kyoto 619-02 JAPAN 
suzuki@at r-la.at r.co.j p 
Abstract 
One of the characteristics of dialogue translation is 
that it is strongly dependent on the situation or the 
communicative goal where the utterance is spoken. 
Moreover, the distance between the language pair 
is great, the possibilities of the transfer diverse and 
it would be difficult to guarantee the equivalence of 
translation. 
In this article, we propose a method of utilizing do- 
main and language specific constraints from the view~ 
point of transfer phase in a dialogue translation sys- 
tem. q~ansfer processing sometimes requires various 
kinds of information: e.g., domain knowledge for el- 
lipsis resolution, language dependent communicative 
structures(focus, theme, rheme, ...). 
Our standpoint is that there are certain language 
specific strategies on producing utterances in commu- 
nication. As steps towards clarifying those stratcgies, 
we first discuss tim issue of typical idiosyncratic gaps 
between two language pairs. Next, to resolve such 
problems we suggest a new framework of incorporat- 
ing domain and language specific knowledge a.s trans- 
fer rides for dialogue translation from the viewpoint 
of transfer phase. Finally we will mention related is- 
sues and further investigation. 
1 Introduction 
The difficulty of translating spoken dialogues between 
two languages is often greater than that of translat- 
ing written text. This is because translation of certain 
expressions in a source language are affected by pre- 
vious context or couununicative goals tinder conllnon 
knowledge for dialogue participauts. If the two lan- 
guages are quite different from each other, possible 
expression candidates increase. This further compli- 
cates tile problems of producing utterances in dia- 
logue translation. 
In the author's recent work\[12\], tbe factors which 
affect lexical choices aud setting the default transla- 
tion considering those factors were discussed for se- 
lecting appropriate verbs. Of course the selection of 
an appropriate verb in given situations is only a small 
part of the great theme of translation in various levels 
of expressiou: e.g., selection of modality correspond 
ing to illocutionary forces, correct choice of articles, 
etc. The problem, however, includes important as- 
pects of linguistic phenomena related to other levels. 
In that article, some examples of relationships be- 
tween Japanese and English verbs were investigated 
using a dialogue corpus translated by human inter- 
preters. Through comparing corresponding verbs in 
Japanese and English, some tendencies were envis- 
aged concerning the characteristics specific to dia- 
logue translation. 
Ill another work\[13/, we discussed the issue of trans- 
lating dialogue utterances through a partitioned se- 
mantic network, part of which is actually lexicalized. 
A method was proposed for managing the semantic 
network, i.e. re-partitioning of the network, which 
yields a more natural translation of utterances. We 
assumed three kinds of partitions: theme vs. theme, 
old vs. new information, and fore- vs. back-ground. 
The prodncing of target utterances starts with pro- 
ceasing along the lines suggested by the source lan- 
guage, and then the given partition can be reparti- 
tioned according to properties of the target language. 
Along this examination, we started to describe ac- 
tual transfer rules for certain kinds of Japanese ut- 
terance that seem to be ditficult to translate directly 
into Euglish, because of lacking substantial words in 
Japanese. This is the background and the motivation 
of the current study. 
Our presuppositions are as follows. 
• The analysis of input utterances are lim- 
ited in the source language. 
Basically, dctinite/indetinite or plural/singular 
markers are not indicated in Japanese explicitly. 
Moreover, various kinds of arbitrary/obligatory 
ellipses of subjects/objects/predicates (occasion- 
ally) are difficult to resolve within sentential in- 
tormation. 
• A translbr processing is perfomned be- 
tween lal~guage dependent feature struc- 
tllres. 
The analysis module outputs a. feature struc- 
tnre, which indicates the relationships between 
predicates and their complements and other in- 
formation including some pragmatic features. 
See tile examllles in the next page. 
AcrEs DE COLING-92, NANTES, 23-28 AOt\]T 1992 7 $ 6 PROC. OF COLING-92, NANTEs, AUG. 23-28, 1992 
2 Idiosyncratic Gaps between 
two Languages 
2.1 The Problem 
For instance, in Japanese a verb may have certain 
number of complements (or the verb has its own va 
lency). Those complements are often eliminated if 
they are already known as old information. When 
honorific expressions (respective, condescend, polite) 
are used, a certain kind of the ellipses (zero-pronouns 
referring the speaker or the hearer) can be resolved 
by using these key expressions in many cases. \[I\] 
However, there are another sort of utterances which 
lack substantial words and are difficult to understand 
individually. As a target of our examinations, we se- 
lected an expression called "dr-expression". Though 
there are a lot of variations for da-expressions\[14\], 
we consider one of their types that has a structure 
: Copula(A, B). It means that A and B is identi- 
cal (in a certain sense). In many cases, we cannot 
translate such a dr-expression without knowing the 
context where the utterance is spoken. 
Our current target domain for dialogue translation 
is "conference registration task". The example ~n- 
tenses referred in this article arc from the doinain. 
Suppose that the following Japanese utterances 
should be translated. 
31: sankaryou wa donoyouni shihavae 
(~e) (topic) (how) (pay) 
bayoi deshou ka. (acceptance) (copula) (inter~'ogation) 
32: saatkaryou wa ginkou-furikomi desu*. 
(~c) (topic) (bank-transit) (copula) 
\[\[SEM \[\[rein S-REQUEST\] 
\[agen !X3\[\[LABEL *SPEAKER*\]\]\] 
\[recp !X2\[\[LABEL *HEARER*\]\]\] 
\[obje \[\[parm !Xg\[\[parm !KS\[\]\] 
\[restr \[\[rein DONOYOUNI-'I\] 
\[entity !X5\]\]\]\]\] 
\[restr \[\[rein BAYOI-SHUULD\] 
\[aspt STAT\] 
\[ages !XT\[\]\] 
\[obje !X8\[\[reln SHIHARAU-I\] 
\[PRAG \[topic \[\[focus !X6\] 
\[topic-mod WA\] 
\[scope !X8\]\]\] 
\[speaker !X3\] 
\[hearer !X2\]\]\] 
The analysis result of J 1 is shown below. 
The translation of J1 into English may be the next 
senteace. 
El: Ho~ can I pay for the (attendance) fee? 
\[\[SEK \[\[reln DA-identical\] 
\[aspt STAT\] 
\[obje !X3\[\[parm !X2\[\]\] 
\[restr \[\[reln SANKARYOU-t\] 
\[entity !X2\]\]\]\]\] 
\[ides \[\[pa_vm !XI\[\]\] 
\[rsstr \[\[reln GINKOU_FURIKDMI-I\] 
\[entity !Xl\]\]\]\]\]\]\] 
\[PRAG \[topic \[\[~oeus !X3\] 
\[topic-rood WA\]\]\] 
\[speaker \[\[LABEL *SPEAKER*\]\]\] 
\[herer \[\[LABEL *REARER*\]\]\]\]\] 
FS-J2. Analysis Result of J2 
Tile t~aturc structures of J 1 and J2 are as FS-J 1 
and FS-J2) 
The literal translation of J2 may be : 
E2: The (attendance) :fee is bank-transfer. 
Of course this English sentence is not acceptable in 
ordinary situations. Accordingly a certain way of 
eomplementation is required. There can be several 
alternatives and it might be difficult to choosc one 
appropriate expression among them. For instance, 
J2 could be translated in various ways if such a eom- 
p\]ementation is performed. 
E2a: The payment should be made by bank-transfer. 
E2b: Please pay by bank-transfer. 
E2e: Would you please pay by bank-transfer? 
\[agen !X73 
\[obje !X6\[\[parm !X4\[\]\] 
\[restr \[\[reln SANKARYOU-a\] 
\[entity !X4\]\]\]\]\] 
\[mann !X9\]\]\]\]333\]\] 
FS-J1. Analysis Result of J1 iThese feature structures are partially modified for explanation. 
AcrEs DE COLING-92, NANTES, 23-28 AOUi' 1992 7 5 7 I)gOC. OF COLING-92, NANTES, AUG. 23-28, 1992 
There are some Japanese expressions (typically 
"da-expreasions") which lack a substantial word(s) 
known for tile dialogue participants. In the previ- 
ous example sentence J2, the substantial words are : 
sankaryou(fee) and ginkou-furikomi(bank-transfer). 
The word sankaryou is the Theme 2 of this utterance 
and it is already known (old information) in the 
previous utterance. And the word "shiharau" does 
not appear in J2, while it appeared in J1..12 is a 
typical da-expression (a kind of copula sentence) in 
Japanesefl\[14\] 
The Rheme 2 of the sentence J2 is obviously ginkou- 
furlkoml (bank-transfer) and it meets the focus of 
the previous question J1. Accordingly the utterance 
J2 satisfies the Grice's maxim of informativity. In 
English, however, we can't say "The fee is hank- 
transfer." Wc have to elaborate the utterance with 
some substantial or functional words. 
Generally such kinds of knowledge for elaboration 
have been provided with domain knowledge which are 
commonly accessed by respective processing modules. 
We propose that the concept of Mel'~uk's lcxieal func- 
tions can be extended for designing special sets of 
domain-dependent lexieal functions. This idea is as 
follows. 
2.2 How to elaborate an elliptical sen- 
tence? 
1,br introducing our methodology, we use the follow 
ing predicate-complements notation (hereafter PS). 
\[\[Pred Predicate-symbol\] 
\[Case-labell Complement i\] 
\[Cas e-label2 Complement2\] 
\[Case-label3 Complement3\] 
• . . \] 
The Japanese utterance J2(I:S-J2) can he illus- 
trated as follows. 
\[\[Pred COPULA\] 
tub\] e SANKARYOU\] 
\[Iden GINKOU_FURIKOMI\] \] 
PS-J2. corresponding to J2 
If we transfer it into English, we have to modify 
tile structure using definite semantic relationships. 
\[\[Prod PAY\] 
\[Agen \[\] \] 
\[Obj e FEE\] 
\[Mamt BANK_TRANSFER\]\] 
PS-E2. tbr transtatloll of J2 
2'fhe contre.st of Theme and llheme is described as follows. 
Theme: Topic, what is communicated about 
Pdleme; Connnenl, what is communic.~ted 
3Tllere are ~. lot of vaa'iations of cla expre~qiolm observed in 
our dialogue corpus. 
The structure PS-E2 could yield two sentences : 
The fee is paid by bank-transfer. 
You pay the fee by bank-transfer. 
lIowever, it is still unnatural because the speaker's 
intention doesn't appear in these sentences. There- 
fore another elaboration is needed for producing a 
more sophisticated utterance. 
The previous utterance J 1 is apparently a question 
that demands a value referred with the interrogative 
"donoyouni" (how). During our pre-transfer phase, 
an Illocutionary Force Type (IFT) of tile given sen- 
tence is extracted.t7\] In this case, the IFT of J1 is 
determined ~us QUESTIONREF. 
So far the assignment of IFTs was decided with 
matching surface expressions. Accordingly tile IFT 
of J2 is first recognized ms INFOI1.M (default IFT), 
becanse the predicate DA-identlcal (colmla) does not 
specify any explicit intention. 
\[ lIFT QUESTIONREF\] 
\[Agen *SPEAKER*\] 
\[Recp *HEARER*\] 
\[Ob2e HOW(!X) \[\[Pred PAY\] 
\[Modal OBLIGATION\] 
\[Obj e SANKARYOU\] 
\[Mann !X\]\]\]\] 
PS-E1. corresponding to E14 
Then, we need supplementary information on PS-E2. 
For example : 
\[\[IFT REQUEST\] 
\[Agen *SPEAKER*\] 
\[Recp *HEARER*\] 
\[Obje \[\[Pred PAY\] 
\[Agen \[\]\] 
tub\] e FEE\] 
\[Mann BANK_TRANSFER\] \] 
PS-E2** for translation of J2 
For rewriting from PS-J2 into PS-E2*, the following 
knowledge should be provided. 
• Elliptical predicate 
• Elliptical relationship(s) 
For translating the above mentioned da- 
expressions we have to cnnlp~eu*ent apl)rol)riate 
substantial words, which are rclatively domain 
dependent. 
• Knowlede on IFT 
Though the system of IF'l? was formerly 
thought to be lauguagc mdependcnt, this as 
smnption turned out too naive. We *low consider 
that therc can be some situations where a tra*ls- 
fer of IFT is required, according to the language 
dependent strategies of producing utterances. 
4The expressions !X indicate that they me (:oreferential 
tags, as ill FSs. 
Acrl!s DE COLING-92, NANTE~S, 23-28 AOt)r 1992 7 5 8 PROC. OF COLING-92, NANTES, AUG. 23-28, 1992 
3 A method of Incorporating 
Domain and Language spe- 
cific Constraints 
When wc limit the target, domain of translation, the 
associate knowledge is also restricted. I\[owcver, wc 
have to be eareflll that even in such a presupposition 
some general knowledge (recta knowledge) is indis- 
pensable to make elaboration tbr elliptical sentences. 
3.1 Domain knowledge 
Within our task domain, we have some kinds of 
enlpirical knowledge of "registration for a confer 
cnee". Wc observe several frequent verbs with specif- 
ically Ciominant eooccurrenee of nouns. Thougb these 
nouns do not always appear explicitly, we carl as- 
sume their existence in the background of utterances. 
l'br example, we can describe preferable candidates 
of complements which have strong relationships with 
a verb "pay" in our target domain, 5 The italic letter 
symbols indicate eoneeptnal categories. 
pay object - (attendance) fee, amoun~ of money 
pay agent dialogue pavticipanl(speaker / imarer) 
pay manner -- bank-transfer, credit card, in cash 
pay - time-destination - date, deadline 
Here we define a set of domain dependent knowl- 
edge after tim notation of lexical fimctions proposed 
by Mel'~uk\[8\]. 
Obje(pay) ::¢. fee 
Agen(pay) -~ participant 
Mann(pay) => bank-transfer 
Tdes(pay) -> dale 
Note that the above descril)tions have a direction. 
The righthand symbols can be conceptual categories. 
Then the relationship between fee and bank-transfer 
can be obtained thro|lgh an indirect path. Such kiuds 
of knowledge can be extracted semi-automatically 
from our dialogue corpus, to'u certain extent. 
,3.2 Language specific strategies of 
producing utterances 
It is natural to consider that there exist eertaiu lan- 
guage specific strategies of producing ut*erances, il> 
tuitively fro|n the prevk)us example dialogue, in 
other words, sonic Iangue*ge dependent coInnnluica- 
tive structures are recognized. Pragmatic constraints 
are derived from the difference of eommunieatiw~ 
strategies with languages. So far, this importance 
has been relatively less regarded eonlpared with other 
linguistic aspects in trauslation.\[2\] 
5Thongh we do llOt ment.iml hm'e, it is signilicant for lllc- 
dicLion of words for sp~l:h-to=Sl)cech translation systellls. 
hi Japane,~ dialogues, tile speaker's intention tends 
to appear in tire sentence final expressions emd it is 
quite often rendered indirectly. 
a2: sankaryou wa ginkou-furikomi desu*. 
This is a neutral da-expression and there is no 
additional information other than indication of the 
method of payment. Ilowew~r, the following exam- 
pies includes some special nuances in their final ex- 
pressions. Both of them are quite natural as responses 
to the qnestion J 1. 
J2a: sankaryou wa ginkou-furikomi 
to naZte oriraasu, (polite da-expression) 
J2b: sankaryou wa ginkou-fur±kom± 
de onegai shimasu. (polite request) 
We think that these Japanese utterances are equiv- 
alent under the given situation (or tile communicative 
goal). In any cases, the method of payment is desig 
nated. The point is how it should bc communicated. 
We can assume the attitnde of the speaker is kept 
consistent in principle. The translation shonld also 
follow this rule, especially in ca.ue of the utterances 
by the secretary of the conference. It couhl affect 
the style of expressions. In faet, we found many 1N- 
FORM utternces really mean indirect REQUEST. 
The indirectness is remarkable in Japanese (using 
da-expressious) and a direet translation can be abrut)t 
or not informative, partly because there seems to be 
no polite copula expressions in Ellglish. Therefore, a 
certain transfer of IFT might be required. 
We have to consider some constraints for al)plying 
such a rule. In this case, the IFq' of the previous 
utterance (Q\[II,',STIONRI'~F) should be known. Ad- 
ditionally the focus of the question is needed. Fur- 
thermore, thesaurus knowledge about predicates and 
complements might I)e referred. 
a.a Incorporating two kinds of con- 
straints 
In our dialogue translation system, a feature struc- 
ture rewriting systeln(llWS) ~; is used \['or transferring 
between I:,l)anese and English.\[3\] An el\]ieient control 
mechanism l;:)r rewriting rules in the IONS is realized 
using ltew~'tling 1;'nvironmenl and Application Cou- 
strainls. 
The Rewriting Environment(RE) is introduced in 
order to control rule apphcation and to maintain rule 
lLexibility alld modularity. Each I~.E is composed of 
a conlbiliatioll of paranleters and their values, which 
provides the condit.ion of rewriting and is dynamically 
changed through the transfer processing. 
~1~ has been exterMed for introdlming a type system &lid a 
pl e\[el'e lice Illed~;~ltlsln. 
ACRES DE COLING-92, NAtal, S, 23-28 not)r 1992 7 5 9 I'ROC. OV COL1NG-92, NANTES, AUO. 23-28, 1992 
Each rewriting rule has its specific application con- 
straiuts(AC). When the ACs unify with the RE in a 
certain phase, the appheation of the rule fires. Thus, 
the transfer process is procedurally performed, ac- 
cording to the REs. This allows individual rules to 
be rather declarative and include few procedural de- 
cisions for lexical choice. 
We implemented the rules that contain domain and 
language specific constraints, extending this RWS. 
Several example rule are shown below. 7 
' Con < Pred > SHIHARAU in 
:PHASE :J-E 
:TYPE :GENERAL 
in = \[\[Pred SHIHARAU\] 
?rest\] 
out = \[\[Pred PAY\] 
?rest\] 
end ~ ' 
Rule-1. Transfer rule for a verb "pay" 
'Con < Pred > COPULA in 
:PHASE :English 
in = \[\[Prsd COPULA\] 
\[Obje ?obje\] 
\[Iden ?iden\] \] 
if Previous.Theme is ?obje 
then set ?pred to pred of Previous.Theme 
Bet parameter :STATUS :COMPLEMENT 
out = \[\[Pred ?pred\] 
\[Obje ?obje\] 
\[Iden ?iden\] \] 
end ~ 
Rude-2. Transfer rule for complementation 
''on < IFT > IHFORM in 
:PHASE :English 
:STATUS :COMPLEMENT 
:PREVIOUS-IFT :QUSTIONREF 
in \[ lIFT INFORM\] 
?rest \] 
if type of luput.Obje. Pred is :action 
then set ?output to \[lIFT REQUEST\] 
?rest \] 
out = ?output 
end'' 
Rule-3. Transfer rule for IFT 
TA concise description for notation of rewriting(tra~mfer) 
rules: The first line of a rule indicates the target feature path 
of rewriting, followed by Application Constraints with combi- 
nations of pm'a~neters and their values; e.g. :Type :Generah 
The patterns in = ... and out = ... indicate the input and the 
output (sub)feature stntcture r~pectively. Some additional 
conditions can be described using if sentences. For referring a 
feature value, a feature path i~a tot~to-down direction can be 
used like as Input.Obje.Pred 
Note that the above mfle~ are partly modified for explanation 
using PSs instead of FSs. 
The explanation for the rules is described as fol- 
lows, though the allowed space precludes the detail. 
The whole transfer process are composed of several 
sub-procedures according to the Rewriting Environ- 
meats designated by the main rule (the top level rule). 
The general framework is as follows. 
First, the rewriting of ellipsis resolution process 
provides the missing zero-pronouns referring the 
speaker or the hearer. Then an Illocutionary Force 
Type is given to ttle top level of tile feature struc- 
ture. After this a kind of normalization is performed 
(so called Japanese-to-Japanese transfer) in order to 
make the (Japanese-to-English) transfer easier. The 
processing of these sub-procedures are regarded as a 
pre-transfer phase. 
The main transfer phase contains 3 sub-procedures 
: idiomatic, general and default. The Rule-1 is an 
example of simple general transfer rules. 
After the main transfer phase, the transfer within 
the English feature structures is performed. The 
Rule-2 and the Rule-3 are applied in this phase. 
Using ttle Rule-2, a Copula predicate structure is 
transferred to another substantial predicate struc- 
ture. When this rule is applied, a local parameter is 
set to the Rewriting Environment. After this, under 
the new RE the transfer of cases (e.g. lden -~ Mann) 
is carried out with another rewriting rule including 
domain knowledge. 
The Rule-3 designates a rewriting of IFT from IN- 
FORM to REQUEST under certain conditions. As 
mentioned in tile previous section, such a transfer 
yields a more natural utterance. 
At present the flexibility of the system is still in- 
sufficient from the viewpoint of context, processing. 
Iiowever, it is possible to control apllying rules by 
means of local parameter setting (like :status :com- 
plement), to a certain extent. 
3.4 Other Examples and the current 
status 
The following examples were described as domain 
and language specific knowledge for translating typ- 
ical "da-expressions" that appear in our target dc~ 
maim The frequency of "da-expressions" iu ATR 
Dialogue Database is as follows. This investigation 
(by Tomokiyo) recognized about 200 different word 
sequences as da-expressions in predicate parts of sen- 
tences in the conference registration dialogues. 
The occurrence of da-expressions: 1,845 
The occurrence of all predicates: 5,200 
(approximat ely ) 
The numbers of sentences and words appeared the 
corpus are respectively 1,666 and 38,258. The rate 
of da-expressions is roughly 35 %. Though tile exact 
percentage of copula da-expressions is not yet calcu- 
lated, it is estimated at 150 ~ 200. Besides, we envis 
age some copula expressions which are *tot included 
in the above investigation, like "to natte orimasu" 
(mentioned in the subsection 3.2). The current task 
ACqES DE COLING-92, NANTES, 23-28 AOUq" 1992 7 6 0 PROC. OF COLING-92, NANTES, AUG. 23-28, 1992 
is to classify the types of copula expressions which re- 
quire certain complementation of substantial words. 
Among them, two typical examples are shown as fol 
lows. 
aa: Tsugi no iinkai wa ikkagetsu go desK. 
E3: The next committee will be held after 
one month. 
J4: XXX-hotoru wa ZZZZ-yen to natte orimasu. 
E4: As for XXX-hotel, it(the room charge) 
costs ZZZZ yen. 
Both of tire above Japanese sentences lack substan- 
tial predicates: e.g. corresponding to "will be held" or 
"costs". For translation of 33, an associative knowl- 
edge(a kind of common sense) is required: 
committee time location ~ be held 
In this example, J 3 is the answer for the question that 
demands the date of the next committee. Whether or 
not a substantial predicate indicating the event lead 
by tire committee and the date(interrogation) appears 
in the previous utterances, that kind of associative 
knowledge (relatively specific to the target domain) 
is applicablE. 
As fbr ,14, an implicit comparison (actually the 
local topic of the dialogue is "the expense of hotel 
rooms") is underlying. In this case, the key to com- 
plemental,on can bE obtained from tile preceding ut- 
terances. It implies that the XXX hoteru with topic 
nmrker "wa" (it seems to be the subject of the sen- 
tence like aa) only designates the field of the copula 
equation. In our current h'amework of analysis of 
sentence by sentence, it is impossible to distinguish 
the difference between J3 and a4. Thereh)re certain 
domain klmwledge is required. For achieving a suit- 
able translation, it should be comlected with the law 
guage specitic constraint of producing (discourse) ut- 
terances. The input PS-J4 (corresponds to the anal- 
ysis result, of ,14) couhl be rewritten into I'S-E4, am 
shown below. 
\[ \[Pred COPULA\] 
\[0bje XXX-hoteru\] 
\[Idea ZZZZ-yen\] \] 
I'S-J4. c.orvesponding to J4 
\[\[Pred COST\] 
\[Obj e \[\] \] 
\[Degree ZZZZ-yen\] 
\[Field XXX-hot el\] \] 
PS-E4. tbr translation of.14 
Am the lexicalization from the P.q 1','4, we could give 
several variations for the cave, Field: as for, *,~ the 
case of, ... if we adopt the generating strategy of the 
prior position of theme (equivalent with the input), 
the result output may be as E4. 
4 Discussion 
4.1 Related Issues 
Ellipsis iv one of the prominent characteristics of 
Japanese spoken dialogue. Concerning the issue of 
identifying Japanese zero pronouns, we have some 
previous works. A theoretical foundation was given 
by Yosbimoto\[15\] and an implementation was per- 
formed by Dohsaka\[1\], in which zero pronouns re- 
ferring dialogue participants (speaker/hearer) are 
identified based on tile usage of honor'tics and the 
speaker's territory within a sentence. As such ellipses 
occnr ahnost obligatorily in dialogue, tim formaliza- 
tion seems to be relatively simple. Of course, the 
resolution of some phenomena requires more complex 
information from the context. 
Kudo\[5\] showed that another kind of ellipsis indi- 
cating objects in the previous sentence could be r~ 
solved with local cohesive knowledge extracted from 
actual corpora. This knowledge consists of pmr tern-- 
plate patterns of successive two se.ntences and enables 
certain eomplementation of elliptical objects. The 
value of iris work is to bave proposed a method of 
senti-automatic acquisition of such knowledge from 
real dialogue corpora.r6\] 
The primary objectiw~ of these approaches was to 
resolve ellipses. Therefore, problems of translation 
tmve not been sutticiently focused. HereaftEr we have 
to pay attention to the insight suggested in the pre 
vious sections. 
As approaches tYom the other viewpoint of knowl 
edge based translation, WE tind sonn! representative 
works in which semantic networks are used tbr rep- 
resenting meaning structure including context. (and 
sometimes world knowledge)inlbrmation, \[10\] \[4\] 
Mel'~uk's Meaning Text Theory is remarkable in corn 
sidering cormnunicative structure of text. '\['Ire al.- 
tempt of knowledge based generating mull,lingual 
text at CMU is also notable, while it does not seem 
to },ave clearly mentioned about tile relationships be- 
tween their interlingua and hmguage specilic conunu- 
nicatiw: strategies. 
Stanwood and Suzuki suggested that the conmnl 
nicative structures somel.ina~s ditfer with languages 
and showed a concept of repartitionmg the given ,let- 
work conliguration. In this study, a semantic network 
is ,~ssumed to have been divided into contrastive par 
titions: Theme vs. RhemE, Ohl- vs. New-information 
etc. An input utterance in the source language is rep- 
resented as a part of the ne.twork. From this start 
point, tile producing a target language utterance is 
processed through repartitionmg the network, if nec- 
essary. \[11\] \[13\] q'his processing model motivated the 
currEnt issue of utilizing dolnain and language specific 
constrailltS ill oar (lialog/ll! I.lanslation 8ysteln. 
ACRES DE COLING-92, NANTES, 23-28 Aot~r 1992 7 6 1 PROC. OF COLING-92, NANTES, AUO. 23-28, 1992 
4.2 Future Directions 
According to Narita\[9\], we can aSsulne two kinds of 
syntactic systems for ally languages. The first is a 
core syntactic structurc that is generally recognized 
as a universal system of syntax. Tbe second syntactic 
structure is rather language dependent and periph- 
eral. Ilowever, this does not mean that tile second 
syntactic system is unimlrortant. Though it is dif- 
ficult to translate into other languages, the second 
syntactic system just reflects the characteristics of a 
certain langnage. It includes many favorite expres- 
sions ill the language. This issue is quite interesting 
also froln tile standpoint of soeioliuguistics and cross 
language eomlnunieation. 
From tile viewpoint of translating dialogues, if an 
exl)ressi(m of a source language is peril)heral and 
there is no corresponding structures in a target lan 
guage, the source struetoure could be transforlned into 
a universal structure before translation. In order 
to perforln this idea, such a transformation should 
be possible to be formalized. 1,'urtherlnore, certain 
implicit (domain- and language-specific) knowledge 
might be needed ill sonic cases. 
Tile target expression in this article, a certain kind 
of "da-expressions', is regarded as a typical second 
syntactic structure described above. Our fnture ef- 
forts will be directed to investigating various struc- 
tures and for refining and extending the methodology 
proposed here. 
5 Conclusion 
In order to lirovide a effective method of translating 
a kind of copnla sentences lacking some substantial 
words, a lnetAlod of utilizing domain and language 
specific constraints are proposed. In this attempt, it 
has been exanfined that both domain knowledge and 
language speeitic strategies of producing utterances 
shouhl be incorporated. The feasibility was shown 
through typical examples and transfer rules, while we 
need still more inw~stigation into those linguistic phe- 
nomena and have to develop tile method of knowledge 
extraction. I,'nrthernlore, the related issues and our 
future directions were discnssed. 
Acknowledgment 
The author thanks Ryo Stanwood for his b,'usic lint 
suggestive idea, which triggered the study, i aln also 
very gratefill to Mrs. Tomoldyo for her cla.ssification 
of da-expressions in Japauesc dialogues. Moreover, 1 
express lily gratitude to l)r. Kureniatsn who gave us 
the opportmlities of our research. And I appreciate 
tile warm encouragement by Mr. Morimuto and the 
effort made by my colleague \[or our spoken language 
translation l)roject. Finally 1 thank Prof. Narita, Mr. 
I(udo and all the people that inspired lilt with novel 
views. 
References 
\[1\] Dohsaka, K. \[dentifyhlg Zer(>Pronouns Refer- 
ring to Persons ill Japanese Dialogue, ATR Tech- 
nical Reporl, TR-I-0117 (1989) 
\[2\] Dorr, l~onnie. Solving Thematic Divergences in 
Maclline Translation, ACL '90 (1990) 
\[3\] lIasegawa, T. The Feature Structure Rewriting 
Systcnl Manual (Revised Version), ATR Techni- 
cal Report, TR-I-0187 (1990) 
\[4\] Kittridge, R. et ak Multi-Lingual Text C, enera- 
tion and tile Me.aning-Text Tlleory, Second In- 
ternational Conference on 7heoreet2cal Issues in 
MacMnc Translalwn of Natural Language(1988) 
\[5\] Kudo, 1. Local Cohesive Knowledge, Proc. of 
COLING-90 (1090) 
\[6\] Kudo, 1. and Knrematsu, A. Context Processing 
Mecbanisnl for a Dialogue Translation System, 
IPSJ, Vo1.33 No.2 (1002) 
\[7\] Kulne, M. et al. A Descriptive Framework of 
Translating Speaker's Meaning, European Chap- 
ter of A CL '89 (1989) 
\[8\] Mel'~.uk 1. et al. A Formal Lexicon in the 
Meaning-Text Theory (or how to do I,exiea with 
Words), Compntalional Linguislics 13 (1987) 
{9} Narita, II. A Contr~.stive Study of English and 
Japanese for Autonlatie Translation of Dialogue, 
Unpublished Report in Osaka University and 
ATR (1991) 
\[10\] Nirenburg, S. et al. Lcxical Realization in Natu- 
ral I~anguage Generation, Second hl.ternat~onal 
Conference on Theorcetical Issues in Machine 
Translatio~ of Natural Language (1988) 
\[11\] Stanwood, IL and Suzuki, M. Some Compnta- 
tional Applications of Lexieal Fimetions, ATR 
Technical Report, TR-L0179 (1990) 
\[12\] Suzuki, M. Lexical Choice in l)ialogue Transla- 
lien, Second Bilateral Workshop o7~ Compuler 
Linguistics, Manchester (1991) 
\[13\] Suzuki, M. Repartitioning of the Semantic Net~ 
work for Translating Dialogue Utterances, Sec- 
ond Japan-Auslralis Joi~l~ Symposium on Nalu- 
ral Language Processing (1991) 
\[14\] Tomokiyo, M. et al. Evahmtion of the outlmt ill 
Japanese to Fngiish MT systems, ATR 7~chnieal 
Repot'l, TR-I 012l (1989) 
\[15\] Yoshimoto, I,i. Identifying Zero l'ronouns in 
Japanese Dialogue, Proc. of COL1NG-88 (1988) 
AcrEs DE COLING-92, NAMXS, 23-28 ^o~rr 1992 7 6 2 PROC. OF COLING-92, NANTES, AUG. 23-28, 1992 
