Kono kusuri-wa itsft.rli sugu kiku (thisj \[medicine\] \[on stonmchache} \[immediately I \[take effect\] (Jl) \[Lit. Fhis medicine takes effect on stomachache mlmediately.l (E':) kiku kusuri itsu sugu kono Figure 1. Example for Dependency Structure Modelin!! A, L, M : Semantic Roles (or Case Markers), A : ABeRIwe, M : Modifier, L : Locative

"To what extent should (or can) we treat semantics of sentences?" is also very crucial to the decision for selecting nr designing tile linguistic model for 107 machine translation. But it might be fairly asserted that the majority of the current "practical" machine translation systems (MT, in short)are structure-bound or syntax-oriented, though almost all of them claim that they are semantics-directed. Semantics are used only for disambiguation and booster in various syntactic processes, but not used for the central engine for transformation, generation and of course not for paragraph understanding (See \[Slocum 1985, pp. 14 ~16\] for a good survey and discussion on this problem; and see also \[Nitta, et al. 1982\] for the discussion on a typical (classical) structure-bound translation mechanism,i.e, local rearrangement method). Here "practical" means "of very large scale commercial systems" or "of the daily usage by open users", but neither "of small scale laboratory systems" nor "of the theory-oriented experimental systems". For structure-bound machine translation systems, both the dependency structure modeling and the phrase structure modeling are very fundamental technical tools. ° \[Lit. Co) ~1~ FHi~a) &l£-~/-lk¢9 I~~/l~tz ~.~C~,~To.\]

Kono kusuri-wa itsft no ue-nl subayai kikime-wo mottedm. S NP (SUB J) (I'RH)) Ni P (On J) Ttlis medicine has an immediate effect

SUB J, PRED, OBJ, ADV : Syntactic Roles.

SUBJ: Subject, PRED: Predicate llead, OBJ: Object. ADV: Adverbial. Figure 2. Example of Phrase Structure Modeling

The semantic network medeling, which is recently regarded as an essential tool for semantic processing for natural languages (See for examples \[SimmOns 1984\].), might also be viewed as a variation of dependency modeling. However modeling problems are not discussed further here. Comparing Figure 1 and Figure 2, note that the dependency structure modeling is more semantics-oriented, logical and abstract, in the sense of having some distance from surface word sequences. 3. Machine Translation vs. lluman Translation

Today's practical machine translation systems (MT, in short) (See for example \[Nagao 1985\] and \[Slocum 1985\].) are essentially structure-bound \]iteral type. The reasons for this somewhat extreme judgement are as follows: (i) The process of MT is always under the strong

control of the structural information extracted

from source sentences; (2) In all the target sentences produced by MT, we

can easily detect the traces of wording and

phrasing of the source sentences; (3) MT is quite indifferent to whether or not the

output translation is preserving the proper

meaning of the original sentence, and what is

worse, MT is incapable of judging whether or

not; (El) (J'l) PIP (ADV) on stonlachache. (4) MT is quite poor at the extra-sentential infor-

mation such as situational information, world

knowledge and common sense which give a very

powerful command of language comprehension.

Now let us see Figure 3. This rather oversimplified figure illustrates the typical process of Japanese-English structure-bound machine translation. Here the analysis and transformation phase are based on the dependency structure modeling (cf. Figure i) and the generation phase is based on the phrase structure modeling (cf. Figure 2) (For further details, see for example \[Nitta, et al. 1984\].). This figure reveals that all the process is bound by the grammatical structure of the source sentence, but not by the meaning of that. Source Sentence: Kono kusuri-wa ~ itsu-ni sugu kiku. Analysis Model Representation:

kiku (TNS: PRESENT ....... SEM: KK, .....)

\[take ffect\] kusuri ( ...... SEM: KS, .....) itst~ (.....) sugu (.....) \[medicine\] \[stomachache\] \[immediately\] Transformation --\[" Maybe some heuristic l'ule, motsu (,...) \[have l

itsu (.....) \[stomachache\] ~ Genelation kono \[this\] kusuri (_...) \[medicinel M! kono \[this\] Phrase Structure Formation: Target Sentence: (El): This medicine has an immediate ffect on stomachache. Figure 3. Simplified Sketch of Machine Translation Process

Thus, the MT can easily perform the literal syntax-directed translation such as 'from (Jl) into (E'I)' (cf. Figure i). But it is very very difficult for MT to produce natural translation which reflects the idiosyncrasy of target language) pre-serving the original meaning. (El) is an example of a natural translation of (J1). In order for MT to produce this (El) from (Jl), it may have to invoke a somewhat sophisticated heuristic rule. In Figure 3, the heuristic rule, HR (KK, KS, ...), can sucessfully indicate the change of predicate which may improve the treatment for the idiosyncrasy of target sentence.

But generally. , the treatment of idiosyncratic gap (IG, in short) such as 'that between (Jl) and (El)'

HR (KK, KS, ...-) suggests \]

the change m the predicate-\] L argument relation. 3 k6ka (...-) \[effect\] M l

sugu (...,.) \[immediate\] is very difficult for MT. It might' be almost impossible to find universal grarmnatical rules to manipulate this kind of gaps, .and what is worse, the appropriake heuristic rules are not always found successfully.

On the other hand, tile human translation (HT, short) is essentially semantics-oriented type or meaning understanding type. 3?he reasons for this judgement are as follows: (i) HT is free from the structure, wording and

phrasing of a source sentence; (2) liT can "create" (rather than "translate") freely

a target sentence from something like an image

(Hagram obtained froln a source sentence (F~gure

4); (Of course the exact structnre of th\].q image

diagram is not yet known);

}IT often refers tile extra-linguistic knowledge

f~uch as the common sense and the culture;

Thus, tIT cart overcome the idiosyncrat:ic gaps

(\]G) freely and unconsciously. (3) (4) \[ ~/ b m" hkc ng \ \ \[ Image Diagram )

\ \[ What tile source \] \] Iarget Sentence(s) . . . . . . . . . . . . . . . . . -~ ~e.Uence(s)nlean(s)/ • . ~%_.~__ J Fiqure 4. Human \]'ra.slation Process Source Sentence(s) . . . . . Citation

\[n order {:o s:imp\]ify the arguments, let us assume Lhat some kind of diagram is to be :\[nvokezd item \]:he understaadiing of the original scntence. 'I/his d:iagram may (or should) be completely free from the sopcrficta\] struclnre such as wording, phras~nF,, subjectobject relation and so on, and may be strengthened and rood:irked l)y varleus exCra--\]iuguistlc knowledge. \] t may be early for hnmalTl to compose the sentetlces such as (J2) arm (E2) from tlu{s kind of :\]aaage din-gram .invoked from (J1). But the sentences su<h as (J'\]), (J'2), (E'l) and (g'2) will never be composed by huma\[\] unc\[(lr tile nolTnla\] eond\].tlons. o ~o) ~.'& f,.ktT~ '¢IcOf, ft'~/~ s J ¢ £tl./a., (J2) Ko,,o ktlsu it.we \[iomll.to 1 lio-ita,,i,.ga sugu tore-ru. \[this\] ~nlellicme I \[if(you) take I \[stomadlacne\] \[soon / \[delmved l o \[l.it. If you Iake this medicine :,ou will sonn be dep\[ived of a stomachache \] # (E'2) (E2\] #(r2)

Now, note that there are b:\[g structural, gaps between (gl) and (\],;\]), and between (,\]2) and (E2), whic.h are tile natura\] ref\]eetJons o\] \]:ingu:istic idiosyncrasy orginated in thc culture, i.c, lhe difference of the way of thJnki.ng. So far we have seen that MT is poor at tile idiesyncrasy treatment and eonverse\]y HT :is good at that. This d:ifference between MT and HT depends on whether or not it has tile ahi\]i.ty c f meaning underst:and:ing.

£n this section, let us examine the idiosyncratic gaps between the two sentences which share the same in meaning but each of which be\].ongs to different language. The reason for comparing tile two sentences is that we cannot: examine the linguistic idiosyncrasy itself. Because, currently, we cannot fix the one abstract neutral meaning without using something like the image diagram (cf. Figure 4) which is not yet elucidated.

In order to examine tlm idiosyncratic gap, we have devised the practical method named "Cross Translation Test (CTT, in short)." The outline of CTT is as follows: \]first, take an appropriate well-written sample sentence written in one language, say English; Let E denote this sample sentence; Secondly, select or make the proper free translation of E in the other \].anguage, say Japanese; \],et J denote this proper free translation; J must preserve tile orlginal meaning of E properly; At the same time, make a literal Lrans\].tion of F, in the same language that ,\] is written in; Let J' denote this literal translation; Lastly, make a Literal translation of J in the same language.that E is written in; Let E' denote tM.s literal translation.

iIere, the "literal" translation means the translation that :\[.s preserving the wording, phras:ing and various senteatial structure of the original (source) sentence as much as possible. Then, eventually we may be able to define (a~d examine) the idiosyncratic gap, \].C, by Figure 5. Iu. other words, we may be able to exam:ine and grasp tile i.di.osyncratie gap by comparl.ng \]:he structure of 1)', and that ot! 1,',', or by comparing that of ji and that of J. ~ ...... E ........................... Iff j' ....-, .... / I '-' - E' ~ . . . . . . . . . . . . . . . ± J - IG: Idiosyncratic Gap I,T: Literal Translation FT: Free Translation E, E': Sentences Written m English J, J': Sentences Written in Japanese Is this paper, we have assumed that:

I,T '= MT and I:T ~ HT, where, MT: Machine Translation, and lfi': Human Translation. Fiyure 5. Illustrative Definitioil of Idiosyncratic Gap Now, note that we (:all assume the re\]atJonshil) ,

i,'\]\] e= MT, ........... . . . . a n d

\]\]'T -& \]HT, where "-" " denoLes "near\]y equal" or "be a\].most equivalent to". Namely, we can assume that the \]itera\] LranslatJon, ILl', which i.s preserving the wording, phrasing and structure of tile source sentence, Js almost e-qu:\[va\].ent to the idealized competence of today's practical structure-bonnd machine trans\]ation, MT. Tim rationale of this assumption has already been discussed in Section 3.

In this paper, the litera\] trans\]ation, \],T (" MT), is performed by tracing the \[)roeedural steps of a virtual machine translation system (VMTS) theoretically. \]1ere, the VMTS is a certain hypothetical system which never models itself upon any actually existing machine translation systems, but which models the general properties of today's practical structurebound machine translation systems.

Now let us observe the gap, IG, by applying CTT to various sample sentences. First, let us take an example with large gaps. Kokkyo-no nagal tonnent-wo nuken~-to yuki-guni de-atta. \[of borderl \[Iongl \[ tunnell (after passing throughl \[snow countryl Iwasl • \[Lit. After passing through the long border runnel, it was the snow country.\] +(E'3) • Tile train came out of the long tunnel into the snow country. (E3) Ressha-wa nagal tonnenl-wo fluke-re ytlki-guni-ni de-ta.

(J3) is taken from the very famous novel "Yukiguni" written by Yasunari Kawabata, and (E3) is taken from the also famous translation by Seidensticker. (E'3) :is the slight modification of \[Nakamura 1973, p.27\] and (J'3) is taken from the same hook. In (E3) the new word "the train \[ressha\]" is supplemented according to the situational understanding of the paragraph including (J3) which may, currently, be possible only for HT.

(J3) is a very typical Japanese sentence possessing the interesting idiosyncrasy, i.e., (J3) has no superficial subject. But in (J3) some definite subject is surely recognized, though unwritten. That is "the eyes of the storyteller", or rather "the eyes of the reader who has already joined the travel to the snow country by the train". So the actual meaning of (J3) can be explained as follows: After I (= the reader who is now experiencing the imaginary travel) passed through the long border tunnel by the train, it was the snow country tha~ I encountered.

Thus (J3) is very successful in recalling the fresh and vivid impression of seeing (also feeling and smelling) suddenly the snow country to the readers. (J3) has a poetic feeling and a lyric appeal in its neat and concise style.

But the English sentence such as (E3) requires the concrete, clearly written subject, "the train \[ = ressha\]" in this case, and this concrete subject requires the verb, "came", and again this verb requires the two locative adverbial phrases, "out of the long tunnel" and "into the snow country". Thus, the original phrase "yuki-guni de-atta. \[ = it was the snow country.\]" in (J3) has completely disappeared in (E3), but the new adverbial phrase "into the snow country \[=yuki-guni-ni\]" appears instead. These drastic changes are made under the strong influence of linguistic idiosyncrasy, and, at the same time, with the effort to preserve the original poetic meaning as much as possible.

Consequently, these changes have invoked a large distant gap, IG between (J3) and (E3). But this gap is indispensable for this translation from (J3) into (E3), HT: (J3) ÷ (E3) , where, \](J3) - (E3) I ~ l( E 3) - (E3) I ~ ~G = large.

One more comment. Note that as a result of this large gap, the literal translation from (J3) into (E'3),

LT: (J3) ÷ (E'3)

where, J(J3) - (E'3)J~J(E'3) - (E'3) I = 0 has failed to preserve the original meaning, i.e., (E'3) is an unacceptable translation which is misleading. Because (E'3) can be interpreted as: After something (=it) finished passing through the long border tunnel, something became (= changed into) the snow country.

However, it is not always the case with idiosyncratic gaps. Lastly, let us now observe the somewhat encouraging example favorable for structure-bound machine translation, MT ("-LT). In the following quadruplet, the gap is not so small but the gapless translation, i.e~, LT ('MT) is acceptable. The following sample sentence (E4), is the news line taken from \[Newsweek, January 18, 1982, p.45\]. #(J'4) (g4) a(E'4) \[Eikyo tile • tie may \]lave saved fi~e flight from a tragic \[kare\] \[kamo-~hire-nail\[ kyfljo-shi-ta \] \[sono\] \[teiki-bin\] \[kara\] \[higeki-teki) repeat performance of tile American Airlines DC-IO crash that Killed 275 \[hanpukul I jikk6 I \[nol \[tsuirakul \[l~oroshi-tal 1275 nin-nol people in Chicago in 1979. \[hito-bitol \[Cllicagc-de\] [1979 hen-nil Kare~'a sono teiki-bin-',~o, 1979 nen-m Chicago-de 275 nin-no hito-bito-wo koroshi-ta (Ed) American-K6ktl-no DC-I(\]-no tsuiraku-no higeki-teki hanpuku-no jikk6-kara

$~J~btc 9)6 b~tX~0 kyfljo-shi-ta kamo-shirenai. kore-ni-yotte kono ki-wa, shisha 275 m¢i-wo dashi-ta 1979 nen-no Chicago-kf~k6-de-no tsuintku-jiko-no higeki-no ni-no-mai-wo sake-eta-to ie-y6. • Lit. It may safely be said that. by this, this airplane could eseapefrom

\[to-ie-you\] \[kore-m-yottel Ikono hikouki\] \[sake-etal \[kara\[ tragic repetition of crash accident of American Airlines \[higeki-tekil \[hanpuku, ni-no~mail \[nol Itsuirakul Ijikol \[nol DC-10 in Chicago Airport in 1979 that produced 275 dead persons. \[Chicago-KflkO-de-no\] \[ 1979 nen-nil \[daslfi-tal \[silishal

The free translation, (Jd) is taken from 1982, p.203\] with slight modifications. For reason of space limitation we have omitted the comments to this example.

Let us see one more example sentence (iS) in order to confirm that the structure-bound MT, which lacks the ability to understand the meaning of source sentences, can produce the barely passable translation, and to try to search for the reason for this. • The soldiers fired at the women and we saw several of them fail. (as) Heishi*taehi-ws on-na.tachi-ni happo-shi-ta soshite \[soldiers\] \[at the womanl \[firedl \[and\] 1~o wareware-wa kare-ra-no s~nin~ga taoreru-no-wo mila

\[we\] \[of them\] \[several\] If all\] Isaw\[ } 0'5)

(E5) is one of the sample sentences in \[Wilks 1975\] where anaphora and references are discussed as the important elements of sentence understanding. As is pointed out by Wilks, a certain extent of understanding is necessary to solve the anaphora and reference problem of the sentence (E5), that is, whether "them" refers "the soldiers" or "the women".

And actually, the structure-bound MT, which cannot understand the meaning of "fired.at" and "fall", may translate "them" into "kare-ra" being indiffercut tO tile anaphora and references. ?in Japanese "kare-ra" denot:es the pronoun of \]male, third person, plural\], and "kanojo-ra" denotes tile pronoun of \[female, third person, plural\], so (,7'5) \].s somewhat misleading translation. Nevertheless, human (i.e. almost all thc~ Japanese readers) can sure\].y understand the sentence (J'5) correctly; that is, they can understand that "kare-ra" (="them") is referring "on-na-tachi" (= "the women") uot "heishi-tachi" (="the soldie-rs"). The reason of this is that the human's brain can understand lille ,leaning of the sentence (J'5) with the support of the colmnon sense like : X fires at Y + Y will severly wounded

+ Y will fall and die, which functions as the compensator for the anaphora and references.

The above example shows that the lack of the anaphoric ability in structure-bound MT may sometimes be compensated by the human-side, which is the encouraging fact for MT.

So far the point we are trying to make clear is that even IG-neglecting MT (= structure-bomld machine translation systems) can generate target sentences that convey the correct meaning of source sentences, when tlre \].att:er are written J.n simple, logical, structures. 5. Conclusions

This paper has dealt with the \].imitations and potentials of structure-bound machine translation (MT) from the standpoint of the idiosyncratic gaps (IG) that exist between Japanese and \]';nglish. Tile conmmrcial machine translation system (MT) curreut\].y on the market: are inept at handling riG since they are still not capable of understanding the nleaning of sentences l:i.ko human translators can, and are thns bound by the ,qyntactic structures of the source sentences. This was pointed out by applying the Cross Translation Test (CTT) to several sample sentences, which brought the performance limitations of structure-bound mach:i.ne translation into sharp relief. But the CTT applications also showed that if the source language sentence Js simple, logical and contains few ambiguities, today's fG-neglectJng machine translation systems are capable of generating acceptable target sentences, sentences that preserve the meaning of the original (source) sentences atrd can be understood.

However ~ source sentences are not always simple, logical and unambiguous. Therefore, to improve the performance of machine trans\]ation systems it will be necessary to develop technology and techniques aimed at rewriting .';ource sentences prior to inputting them into systems, and at formalizing (norma\]izing) and control.ling source sentence preparation. One move in this direction in recent years has had to do with tile source language itself. Research has been steadily advancing in the area of Sub\].anguage Theory. Sublanguages are more regulated and controlled than everyday humml languages, and therefore make it easier to create simple, logical sentences that are re\].atively free of ambiguities. Some examples of sublanguage theories currently under study are "sublanguage" \[Kittredge and Lehrberger 1982\]~"controlled language" \[Nagao \].98311 and "normalized language" \[Yoshida \]984\]. The aim of these sublanguage theories is to assign certain rules arld restrict:lens to \]:he everyday human \]anguagea we use to trausmXt and explain information, improving the accuracy of parsing operations necessary \]for nlachJ.ne processJ.ng~ aud enhancing human understanding. Some examples of the \].ingnJstic rules and restrictions envisioned by the sublanguage theories are rules governing the creation of lexicons \[Kigtredge and Lehrberger 1982\], rules governing the use of function words related to the log:tca\] construction of sentences \[Yoshida 1984\] and ru\]es governirlg the expression of son\]cut\]el dependencies \[Nagao 1.983\] . References Eikyo \[Nihon-Eigo-Ky~Jku-Ky$kai\] (eds.) (1982), '2

Ky~t Jitsuy$ Ergo Ky~hon' ('2nd Class Practical\]

English Textbook' ) , N:/hon-Eigo-Ky$iku-KySkai,

Tokyo, 1982 pp.202-203 (in Japanese). Kittredge, Richard and J. Lehrberger (eds.) (1982),

'8ublanguage: Studies of Language in Restricted

Semantic Domains', Walter de Gruyter, Berlin,

New York, 1982. Nagao, Makoto (1983), 'Selgen-Gengo-no Kokoromi' ('A

Trial in Control.led I,anguage'), in ShJzen-

(~n_\]39-Shori-G"i j nt su _S3£m~iun h Yok-~--~l$,'=fn f o r -

mat\]on Processing Society of Japan, Tokyo, 1983

pp. 91-99 (in Japanese). Nagao, Makoto (\]985), 'Kikai-Ilon-yaku-wa Doko-made

Kan$--ka' ('To What Extent Can Machine Trans-

late?'), K ag~u., lwanami, Tokyo, vol. 54 no.9,

1985, pp. 99-107 (in Japanese). Nakamura, Yasuo (\]973), 'ilon-yaku-no Gijutsu' ('Tech-

niques for Translation'), Ch~-k$-Shinsho 345,

Ch(~8)-K~ron-Sha, Tokyo, 1973 (in Japanese). Newsweek (\]982), 'Newsweek' January, 18, 1982 p.45. Nitta, Yoshihiko, et al. (1982), 'A Heuristic Approach

to F, ug\]ish-into-Japanese Machine Translation',

in J. Horocky (ed). Prec. (\]OLING82:__~t Prgg ~

\[Proceedings ............................................ of the 9th Internatim:a\] Conference

on Computationa\] Lingnistics\]. ~'Iorth-Ilolland

Publishing Company, 1982, pp.283-288. NJtta, Yoshihiko, et al. (1984), '£ Proper Treatment

of Syntax and Semantics in Machine Translation',

:in Prec. COLING 84 (at Stanford_) \]Proceed\]ntis

o£ the-lOi:l~' International Conference on Coln--

putatdona\] Lingudstics\], Association for Com-

l~-UTatl~ffa\]7\]Ti-ngui.~ics, 1.984, pp. 159-166. Simmons, Rohert F. (1984), 'Computations from the

English', Prentice-Hall, Englewood C\].Jffs, New

Jersey, 1984. S\]ocum, Jonathan (1985), 'Machine Translation: Its

llistory, Current Status and Future Prospects'

Cqnli~utational LJ_nguist.ies, vol. 1.1 , no.l, 1984,

pp. 1-17. Wilks, Yorick (1975), 'An Intelligent Analyzer atrd

Understander of EnglJ.sh', Communications of the

ACM, vo\].18, no.5, I!)75, iIp.264-274. Winograd , Terry (1983), 'Language as a Cognitive

Process: vol. I : Syntax', Addison-Wesley, Menlo

Park, Calif. 1983. Yoshida, Sh$ (1984), 'Nihongo-no Kikakuka-nJ-kansuru

KisotekJ Kenky~,' ('Basic Study on the NornmlJza-

tion of Japanese Language'), Sh\]wa 58-.non-do

Kagaku Kenky~-Hi IIojokin Ippan-Kenky-~-TB)

Kenkyu-Seika Hokoku-Sho (Research Result Report

on the General\[ Study (B) Sponsored by tile Sh~wa

58 Fund for Science Researc\]1) Kyushu University,

Kyushu, \]-98~n-Japat~e)-7