A METHOD OF TRANSLATING ENGLISH DELEXICAL STRUCTURES INTO JAPANESE
Hideki TANAKA~, Tcruaki AIZAWA, Yeun-Bae KIM and Nobuko HATADA Science and Technical Research Labs. of NHK ld0-11, Kinuta, Setagaya, Tokyo, JAPAN ttanakah@ strl.nhk.or.jp (junet)
1 Introduction rioratcs the quality of the translation. For ex- Common verbs such as "take," "have," ample, "make an address" becomes "enzetsu wo "make," and "give" appear frequently in Eng- tsukuru," which meaTts "create an address." lish. These verbs quite often constitute verb - There are two possible ways of translating a deverbal noun structures such as "make an ad- DS. The first is the idiosyncratic approach, list- dress," "give an answer" and "take an approach." ing all the DS's with their Japanese translations The verbs in the structure arc ahnost devoid of in a lexical system. This approach, however, lexical meaning but bear syntactic information suffers fi'om several shortcomings: such as tense, number and person; the deverbal (1) The DS's are numerous and hard to list noun carrying the lexical meaning. The verbs, exhaustively: some DS's allow pttssivization in this case, are called "dclexical verbs" and some deverbal nouns can be modified [Collins] or "light verbs," [Live] which refers to by quantifiers, adjectives and so on. This their lexical emptiness. doubles and triples the number of possible In this paper, we call such verbs Delexical DS combinations. Verbs (DV) and a "DV + deverbal noun struc- (2) This direct method is unable to infer the ture" a Delexical Structure (DS) Iollowing the translation of a DS undefined in a lexicon. examples ol' [Collins]. The frequency of these (3) The use of this approach increases the verbs in actual text can be seen, for example, in nulnber of lexical entries, making lexical the COLLINS COBUILD ENGLISH DIC- TIONARY, where the pmagraph on the entry management difficult. "take" states: "The most frequent use of take is in Another approach is to synthesize the transla- expressions where it does not have a very dis- tion of a DS using the word sense of each com- tinct meaning of its" own, but where most of the ponent with syntactic and semantic rules. The meaning is in the noun that follows it..." attractive part of this "synthetic approach" is that We have been developing an English to Japa- it does not suffer from the problems mentioned nese machine translation system for news broad- above. The "ntonosenty approach" proposed in casts since 1989 [Aizawa] [Tanaka]. The precise [Ruhl] can be viewed as the extreme manifesta- translation of DS's in news texts is of great im- tion of the synthetic approach. A recent lexical portance since they are quite frequent here. We framework [Bograev] proposes to generate the counted the number of "take" + "noun" colloca- word sense instead of listing them exhaustively tions (as verb + object) in 21 months' worth of in a lexicon, which is similar to the synthetic AP texts using the parser of the machine transla- approach. tion system. "Take" collocated with 2,188 dif- However, fl'om a practical viewpoint, not MI ferent nouns a total of 20,271 times. Of the col- DS's can bc translated by this approach as the locating nouns, 87 deverbal nouns were found necessary rules have not yet been factored out. out of the 119 deverbal nouns listed in [Live], We propose the use of a DS translation method comprising about 28% (5,726) of all occur- based mainly on synthesis, and the employment rences. This figure strongly supports the state- of all idiosyncratic approach where synthesis is ment in the Collins Dictionary. difficult. To do this, DS's were categorized into Failures in DS translation typically result three groups, called type-l, type-2, and idio- from producing the primary sense instead of the matic DS. The first two groups are translated by delexical sense of the DV, which greatly dete- the synthetic method and the last group is trans-
ACRESDE COLING-92, NANTES.23-28 AOt~r 1992 5 6 7 PROC. OF COLING-92, NA~CrES,AUG. 23-28, 1992 lated by an idiosyncratic approach which can and euphemistic impression [Yamanashi] than hopefully be integrated into the former part as the corresponding verb expressions and used research reveals the underlying rules. This frequently in modern English [Live]. method should provide clear distinctions be- Another feature of a DS is that it gives the tween idiomatic and synthesizable DS's through impression that the event is brief.' the use of a set of rules, which would facilitate She gave a scream. (brief) the management of lexical systems. She screamed. (not brief) The translation rules are quite simple tbr the [Collins] following reasons: Some syntactic operations are possible on (1) English DS~s have Japanese equivalents DS's. For example, adjectives and possessive in many cases, and some parallels can be pronouns can modify the deverbal noun, and seen between them. some DS's can be used in passive forms. (2) Many Japanese "verbal nouns" take the In this paper, determiners are mainly consid- form of "sahen-meishi," which become ered as the modifier of the deverbal noun since verbs by simply adding "suru" to the tail. they are the basic modifier of the DS's. The verbs "take," "make," "have," and "give" are However, some DS's require translation in a studied since these are the most common DV's. passive sense. The conditions were factored out As a result, we are able to obtain a DS in the through semantic consideration, and were inte- following form: grated into the translation rules. The rules were implemented in the machine DV + (det) + NV (1) translation system and AP news texts were 0: optional translated appropriately, thus proving the feasi- DV: Delexieal Verb (take, make, bility of this method. have, give) det: a, an, the 2 Delexical Structures NV: Deverbal noun This section introduces some definitions and characteristics of DS's. ADS basically takes the 3 Categorization of DS's fo ml: The previous section mentioned that a DS Delexieal Verb + (determiner) + Deverbal chooses a single sense of the corresponding Noun verb. An example is shown in figure 1. 0: optional DS as in "give an answer," "make an array," "take an approach," "have a battle," "take ad- C give all answer) vice," and "make an announcement." Verbs such as "put," "cast," "pay," "lay," and "throw" answer (verb) are also known to work as delexical verbs [Live]. These are commonly-used transitive ~ 1 say, write or do something in verbs. response to somebody DS's have some interesting semantic charac- teristics. A verb can be substituted for a DS in 2 be suitable for many cases, but the substitution is one way: thus I "step" for "take a step" but not "take a move" for Figure 1 Word sense selection "move" [Live]. The verb-substitutable DS's choose one of possibly several word senses of Since a deverbal noun in a DS possesses the corresponding verb except when they are some nominal sense of the matching verb, we metaphoric. The selection is controlled by the can assume that a DS chooses a single sense of co-occuring DV. Thus the same deverbal noun the deverbal noun and verbalizes it as shown in can form DS's which represent different mean- figure 2. ings. We term this kind of DS an ordinary DS. A make an order ==> request for supply DS which cannot select any sense for a deverbal of goods noun is termed an idiomatic DS. give an order ==> command An ordinary DS can further be categorized The DS's give a greater colloquial [Konishi] into a type-1 DS or a type-2 DS. ADS which verbalizes the primary meaning of a deverbal
Ac'rEsDE COLING-92, NAN'I'ES,23-28 AOt~T 1992 5 6 8 PROC. OF COLING-92. NANTES.AO~. 23-28, 1992 DS pronouns, the translation mechanism should also allow the same degree of flexibility. ~ (give an answe r ) 4.3 Type-2 DS This type can be translated iu approximately answer(noun) the same way as type-l. However, one differ- ence is that this type verbalizes the peripheral 1 thing said, written or verbalization Japanese translation in the lexicon. Since insuf- done.as a response or ficient knowledge has been accrued on what reaction ~ makes the DS select the peripheral meaning, the collocation of DV's and deverbal nouns should I 2 solution be specified in a lexicon. As this type also per- mits some modification and allows passive Figure 2 Word sense selection transfommtion, it should be translated by syn- and verbalization thetic approach. noun is a type-1 DS and one which verbalizes 4.4 Idiomatic DS translation other meanings is a type-2 DS. The whole cate- This type does not verbalize any translation gorization is summarized below. of the deverbal noun and is "frozen" in that the type-1 DS deverbal nouns are rarely modified by adjectives and the DS does not allow passive transforma- ordinary DS / tion. This type of DS must be listed in a lexicon F L type-2 DS in its full form with a proper translation. The DS L idiomatic DS implied definition of an idiomatic DS here is (1) it does not represent any word sense of individ- Figure 3 Categorization of DS's ual components and (2) it resists any form of transformation. 4 Translation strategy This section describes the basic strategy of 5 Passivization rules DS translation. A concrete implementation will Some DS's can be substituted with a match- be described in section 6. ing passive verb as shown in [Live], so that 4.1 Parallels between Japanese and English "take comfort" and "take counsel" can be para- The DS categorization in section three was phrased as "be comt'orted" and "be counseled," based on the selection of word senses for an respectively. Though the number of such cases English deverbal nouns. In many cases, we can does not seem to be large, it is important to assume parallels between English and Japanese: translate them passively, otherwise the meaning (1) The word sense of an English deverbal of the translation is totally reversed. Most DS's mmn can be translated into a Japanese word which can be substituted by passive verbs should be translated in a passive sense in Japa- sense of a deverbal noun. nese because of the parallels between the two (2) The verbalization of an English word languages. "Take comfort" and "take counsel" sense in a deverbal noun and the should thus be translated "nagusame rareru" and verbalization of the corresponding Japanese "jyogen wo sareru." equivalent can represent the same meaning. In this section, some rules tot passivizafion 4.2 Type-1 DS arc obtained by analyzing the [Live] table. Taking advantage of these parallels, the Conditions for these rules will be sought in the translation of a type-1 DS can be achieved by syntactic and semantic characteristics of botb verbalizing the core Japanese translation of the the DV and the deverbal noun. deverbal noun in a translation lexicon. Table 1 is extracted from the table in [Live] As we can see from the table in [Live], type-I listing the combinations requiring passive sub- DS's are predominant. This justifies the use of stitution. type-1 as the default translation of unknown 5.1 Condition for DV DS's. "Take" and "have" have about 10 passive As most of this type can be used passively, verb substitutable DS's, whereas "make" and and since deverbal nouns of this type are some- "give" have only one. This suggests that condi- times modified by adjectives and possessive tions depend on the contrasting characteristics
ACRESDE COLING-92. NANTES,23-28 AO~r 1992 5 6 9 PROC. oI~ COLING-92, NANTES,AUO. 23-28, 1992 of "take," "have" and "make," "give." "Take" As the deverbal nouns in Table 1 are derived and "have" have a common characteristic: the from transitive verbs, we can picture a subject + action moves from the object to the subject. verb + object structure for them. With "make" and "give," on the other hand, X + Vt + Y (3) the action moves from the subject to the object. X: Subject The characteristics of the take-have type Vt: Original Transitive Verb of the seems to be the DV condition. Actually, we can NV see some hyponyms of take-have type forming Y: Object passive verb substitutable DS's: Passive verb substitution occurs when the I received encouragement from my Sub in (2) agrees with Y in (3). friends. -~> I was encouraged by my fiiends. He gave advice. He took advice. A ~ ..... Table 1 Combinations constituting passive X advised Y X advised Y verb substitutable DS's [Live] 1 DV's take Figure 5 Agreement between Y and Sub comfort, counsel, hint, shape, advice, The conditions which make Y in (3) agree encouragement, examination, offense, with Sub in (2) can be found in the syntactic and 31easure, satisfaction semantic characteristics of the verb Vt. DV's have "Transitive" is one of the conditions men- 9romise, reward, approval, disturbance, tioned previously. This is not, however, suffi- diversion, examination, excuse, impression, cient as some transitive verb derivative nouns do not require passive verb substitution. ~leasure, reversal, satisfaction Some examples are shown in the table 2. DV's make Table 2 Combinations which constitute excuse active verb substitutable DS's [Live] t DV's give excuse DV's take approach, average, bet, command, count, dare, dislike, etc. DV's have approach, array, battle, bet, catch, change, Make.Give type claim, command, debate, desire, dislike, display, dispute, doubt, dread, etc. The deverbal nouns in Table 1 are a confined subset of Table 2. A common structure seen in the verbs in Table 1 is: o:olTake-Have type X
AcrEs DE COLING-92. NANTES. 23-28 Ao~q" 1992 5 7 0 PROC. ol: COLING-92. NANTES.AOO. 23-28. 1992 [Yamanashi] formed passive verb substitutable result of the action implied by the verb. DS's. Condition (3) can be discussed from the exempt viewpoint of transitivity. Since the effect on the Ted had an exemption from military service object is strong, strong transitivity occurs because of his poor health. [Ikegami]. However, the result of the effect is => Ted was exempted fl'om... not always certain and is up to the will of the permit object
ACRESDE COLING-92, NANTES,23-28 ^o(;r 1992 5 7 1 Pn(~c, ol; COLING-92, NAN'rraS,AUG. 23-28. 1992 Verb marker fieM ...... (make an exceptioq !:r...... ~ take exception to J translation passive
i i Oeverbal Noun [wo sareru p entry
translation marker field Deverbal Noun type verb passive marker field jyogai typed translation t e verb assive (excluding) chuukoku type-1 p ..) igi moushitate type-2 take (advice) (to object to) kankoku (authority's advice) shinsatsu Figure6 Lexicaldescription (medical examination) for a deverbal noun simply by rendering the translations of "make" and "give" as "wo suru," and of "have" and ftakeadvicex, passive "take" as both "wo sareru" and "wo sum," one of ~"~ (chuukoku wo Sareru "~ which will be selected. ~k chuukoku wo suru j An example lexical description for passiviza- default and active ~- tion and verbalization are shown in figure 7. "Take advice" and "give advice" are trans- Figure 7 Lexical description for passivization lated in passive and active voice respectively. and verbalization The sentences which were translated cor- 7 Translation Experiment rectly by both of the dictionaries were evaluated The DS's contained in the AP texts were "Equal." The "Correct Rate," including "Equal" translated using this mechanism. Two dictionar- responses, reached nearly 100%. ies were compiled to test the effects of the Some of the results are shown below. mechanism. In one of the dictionaries, neces- Upper translation is with the DS tranMation mechanism sary markers were ascribed to 302 deverbal Improved Examples nouns - appearing in [Live] - and 4 DV's (take, Ca~ l: But we think they'll make a right choice eventu- make, have, give ). The other dictionary had no ally." such modifications. Lz3' L, "~LG75~),,3, 2I- bv~'~-5(sentaku All the March 1991 AP texts which contained both the DV's and the deverbal nouns mentioned L;O" b, ~fL;97)¢~aiJr,~, ~(sentaku above were selected by a simple filtering of the wo tsukuru) ~ ~4" ~j: tL~..~o Case 2: Unlike Soviets, Americans seem to have made text. The filtered text was translated using the progress against the stigma of alcoholism, Zubkov said. two dictionaries and the results were compared ,; ~':L b L ~ o T 7 3 ,) :# AIJ. y Jv ~ - )t,~:~e) by human assessment. Sentences not having a ~.~:~S_~ b'V~(shinpo wo shita) ~ 5 "~oTa "verb - object" relationship between the DV and L, Zubkov ;~bi-'~7~2o deverbal noun were rejected. The results are shown in Table 4. Table 4 Experimental Results "~ ~ ~ ~, Zubkov ;~ii" o/~.o Case 3: After his Toyota Celica GT4 took the lead, he DV give have make take held off challenges from a trio of Lancia Delta lntegrales. Improved 35 115 349 302 '~© Toyota Celica GT4 h ~ ') -- b"~ b L:(fiido wo shita) Equal 13 44 0 2 ~, ~I~ Lancia Delta Integrales © 1" ~) ff ~" ¢9 ~[~ Made Worse 4 9 2 2 ~t~e) Toyota Celica GT4 7~~(namari wo totta) Total 52 168 351 306 ~, ,~12 Lancia Delta [ntegmles 60 b ~) ff7)~ ¢9~i~ Correct Rate 92% 95% 99% 99%
ACRESDE COLING-92, NANTES, 23-28 AO~r 1992 5 7 2 PROC. OF COLING-92. NANTES, AUG. 23-28. 1992 Case 4: However, they took some encouragement from To achieve better results, precise verbaliza- the fact thai Februmy's decline was much smaller than the tions of Japanese nouns are required, though a previous setbacks. simpler method worked well due to the ability to express many "verbal nouns" as "sahen-meishi" in Japanese. ~'~1/~!:~ ~ ;~Lf~:(gekirei wo sareta). A voice control rule was developed based on the "flow of action" of delexical verbs and the :t~j ~- ~: .9 ~,=(shourei wo totta)o "transitivity" of deverbal nouns (the original Non-improved Example verb). Other delexical verbs such as, "put," Donovan "Razor" Ruddock gives the impression he "cast," "receive," etc, can also be categorized by doesn't live and breathe boxing. "flow of action" and can be integrated into tbe Donovan Razor Ruddock ~:~'~7~t L.~;~~ same framework to cover a wider range of DS (inshou wo shire), "~ [~ ~C ~J~ ~/5" ?/~" ~: I['~:1~-~ ~ o translations. Donovan Razor Ruddock t~,~ ~ L 7~,;v~[~ll~ ~ ~j. ~. ~C (inshou wo ataete), @ [~ ~C ~IZ~" -.~ ?/~/-~ 11'~'-1~'9~ 7~) o The role of the other DS components, ar- ticles, possessive pronouns, and prepositions Translations without the DS translation should also be explored. mechanism produced the origiual word sense of The differences between "make" and "give," the DV, resulting in awkward translations. and "take" and "have" has been neglected in this In Case 3, the DS correctly translated " take rule. However the differences between "take" the lead" using "riido wo suru (go ahead of)" and "have" in certain contexts have been pointed while the original system produced "namari wo out by [Wierzbicka]. It would be interesting to toru (pick up a piece of the metal lead)." The explore these differences in the context of trans- revised mechanism thus successfully distin- lation. guished the ambiguity in the word sense of [References] "lead." In Case 4, the passive verb substitutable [Ai:,~twa] T. Aizawaet al.: "A Machine Transla- DS "took some encouragement," in which the tion System for Foreign News in Satellite deverbal noun is modified by "some," was suc- Broadcasting," Proc. of Coling-90 Vol. 3 pp. cessfully translated passively into Japanese. 308-310, (1990) In the example which was made worse, "give [Boguraev] B. Boguraev and J. Pustejovsky: an impression" was translated into "inshou wo "Lexical Ambiguity and The Role of Knowl- shite," which is worse than "inshou wo ataete." edge Representation in Lexicon Design," This is due to the simple verbalization mecha- Proc. of Coling-90 Vol. 2 pp. 36-41, (I 990) nism which attaches "suru" to all nouns backed [Collins] "Collins Cobuild English Grammar," up by heuristics. The transitive verb form of the Collins (1990) Japanese word "inshou (impression)" is "inshou [lkegami] Y. Ikegami: "Suru to Naru no Gen- zukeru." Further elaboration of the verbalization gogaku," Taishuukan, (1981 ) mechanism, such as the use of a noun/verb con- [Kizuka] H. Kizuka: "A Dictionary of Verb - version table, will solve this problem. Noun Collocations," The Japan Times, The quality of the translation was improved (1987) remarkably, though the translation mechanism [Konishil Y. Konishi: "A Dictionary of English was quite simple. Word Grammar on Verbs," Kenkyuusha, (1980) 8 Conclusions [Live] A.H. Live: "The Take - Have Phrasal in A method has been proposed for translating English," Linguistics Vol. 95 pp. 31-50, English delexical verb structures into Japanese. (1973) Since this method involves rule-based syn- [Ruhl] C. Ruhl: "On Monosemy," SUNY, thesis, transformed and modified DS's were (1989) flexibly translated. The conditions which re- [Tanaka] H. Tanaka: "The MT User Experi- quire a DS "passive verb substitution" were fac- ence: Panel Discussion," Proc. of the MT tored out and included in the translation rule. Summit I11 pp. 121-126, (1991) The feasibility of this method was tested by [Wierzbicka] A. Wierzbicka: "Why Can You translating AP news texts, which sbowed a re- Have a Drink While *You Can't Have an Eat?," Language 58 pp. 753-799, (1982) markable improvement in the quality of transla- [Yamanashi] M. Yamanashi: "Hatsuwa Koui," tion, with a correct rate of nearly 100%. Taishuukan, (1986)
AC~'Es DE COLING-92, NANTES,23-28 Ao~q" 1992 S 7 3 I'ROC. OF COL1NG-92, NANTES, AUG. 23-28, 1992