Machine Translation without a source text

Harold L. SOMERS, Jun-ichi TSUJII and Danny JONES

Centre for Computational Linguistics UMIST, PO Box 88 Manchester M60 1QD, England

Abstract by introducing 'understanding' based on 'domain specific knowledge' (as in the Tiffs lmper concerns an approach to Machine Translation 'sublanguage' approach - cf. Kosaka et al. whieJJ differs from the typical 'standard' approaches crucially 1988, Lehrberger & Bourbeau 1988). This in.that it does not rely on the prior existence of a source text course of research would be inevitable if we as a basis of the translation. Our approach can be character- were to confine ourselves to translation of ised as an 'intelligent secretary with knowledge of the prepared texts which already exist before foreign language', which helps monolingual users to formu- translation. In such cases, we have to recover late the desired target-language text in the context of a (key- from text itself or by using extra 'knowledge', board) dialogue translation systems. such implicit information which is necessary for Keywords: Machine translation; natural formulating target expressions. language interface; dialogue However, we can imagine a quite different course of research for developing a different type of MT system, i.e. an 'expert' system Introduction which can play the role of an 'intelligeut Machine Translation (M'f) or natural secretm-y with knowledge of the toreign lang~lge translation in general is a typical language'. Such a system does not require the example of the 'under-constrained' problems user (the writer) to prepare full source texts in which we often encounter in the field of advance. It slarts from rough sketches of what artificial intelligence 1. That is to .say, the same the writer wants to say and gathers the 'messages' can and should be translated information necessary for formulating target differently depending on the surrounding con- texts by asking the writer questions, because the texts (where and when they are used), and on wdtor is the person who really intends to the Sl~eakers' intention (what they really want to communicate and has a clear idea about what express) etc. It is all too often the case that this s/he wants to say. We can get much richer information, which is neces~ry for the selection information through such interactions than in the of the appropriate overall target text structure, is usual written text translation by professional not ntade explicit in source texts prepared for translators. Through interaction, we can get translation. The author of the source text natur- information concemed with, for example, the ally follows the 'rules' of the source language in user's intention which is not explicitly expressed preparation of source texts and assumes that the in the 'text' to translate but which is nonetheless factors which will affect the selection of target necessary for producing quality target texts. expressions are self-evident. This sort of system is different from the MT systems developed so far or being widely promoted 'Translator's Workbench' idea developed have been trying to compensate this (e.g. Kay 1980, Melby 1982), the main aims of genuine property of language translation by which are to help translators to translate texts. extending the units of translation from sentences In this scenario, both the system and the user to texts (e.g. Rothkegel 1986, Weber 1987) or have knowledge about both source and target language, and it is sometimes difficult to see t The authors would like to acknowledge the contribution where the most appropriate division of labour to this work of the other members of the project team: Bill should occur: indeed, there is sometimes a Black, Jeremy Carroll, Anna Gianetti, Makoto Hirai, Natsuko conflict between what the system offers the Holden, John Phillips and Kenji Yoshimura. translator-user, and what the user already

l 271 knows, or between the extent to which the research programme into automatic system or the user should take the initiative, interpretation between English and Japanese of which might differ from occasion to occasion. telephone conversations. As such it is oriented On the other hand, in the proposed expert towards translation of dialogues. One approach system scenario, the partition of knowledge is to dialogue translation has been the clear: the system knows mainly about 'phrasebook' approach of Steer & Stentiford translation, the writer knows only about the (1989). In this speech translation prototype desired communicative content of the message. system, set phrases are stored, as in a There is no conflict between what the system holidaymaker's phrasebook; they are retrieved assumes to be the extent of the writer's (the by the fairly crude, though effective, technique user's) knowledge, nor in the writer's of recognising keywords in a particular order in expectations. In this respect we are following the input speech signal. The main disadvantage the line taken by Johnson & Whitelock (1987), of this system is its inflexibility: if the phrase and the work here at UMIST on the ENtran you want is not in the phrasebook, you cannot project (Whitelock et al. 1986, Wood & say anything. Chandler 1988) developing an MT system for a In the research programme to be reported monolingual user. here, we are not concerned with speech MT systems so far have been developed processing per se, and we assume the context of based on the implicit assumption that source an on-line keyboard conversation function such texts contain all (or almost all) the information as talk in UNIXTM (cf. Miike et al. 1988). It has necessary for translation. We take as a starting been found that keyboard conversations have the point that this assumption is not necessarily same fundamental features as telephone true, especially when we consider pairs of conversations, notwithstanding the obvious unrelated languages where cultural as well as differences between written and spoken linguistic differences contribute to this problem. language (Arita et al. 1987, Iida 1987). Notice that the concept of 'source text' in the Furthermore, we restrict ourselves to goal- above is quite different from that in the normal oriented dialogues, i.e. dialogues where one context of MT. That is, we do not have a source participant is seeking information from the text to translate as such, but instead, the user other: our experimental domain is dialogues for has his/her communicative goals and the a conference registration and hotel reservation translation system can help to formulate the system. most appropriate target linguistic forms by When such conversations are subjected to the gathering information necessary to accomplish additional distortion of being transmitted via a these goals through 'clarification dialogues'. traditional MT system, several further problems It could be argued that this generation of a accrue, as the talk experiment mentioned above target text on the basis of something other than showed, notably when mistranslation occurs. a source text is not 'real translation'. Such an The problem of human-machine interaction in argument might derive from an overly the specific area of clarification dialogues for traditional view of translation where a translator MT must be studied. The need to incorporate gets some text (say, in the post) and sits at a different types of clarification dialogue has desk with a bilingual dictionary and translates general implications for the question of system 'blind' i.e. with no actual knowledge of the architectures for interactive MT systems. This writer's intentions, goals, etc. There is a sense aspect is discussed in detail below. in which second generation MT systems simply In the above scenario, the system tries to reflect this scenario of a translator. Of course, gather information necessary for formulating the best translations are done by a translator target texts through interactions. This means the who can ask the original author "What did you system formulates target texts by adding mean when you said...?"; by the same token we information to 'source texts' (in the believe we can build a better translation system conventional sense). We can extend this idea if we can elicit such information from the further. In the extreme case, we can hnagine a originator of the 'text' at the time of 'writing'. system which has stereotypical target texts in certain restricted domains (e.g. business General background to the research correspondences in specific areas), retrieves This research is undertaken in the context of appropriate texts through dialogues with users the more general activities of the Japanese ATR and reformulates them to fulfill the specific

272 2 requi~-ements expressed by users. In this First, in I)MT there are several different scenario, the MT sysmm becomes a kind of types of dialogues, any of which may start up or multilingual text generation system and adds a be resolved at any given time: these dialogues lot of inlormation not contained in the 'source include text' at all. This idea has becn investigated here (a) usermser object-level dialogues at UMIST in the context of a research (b) user-user metadevel dialogues (e.g. in programme for British Telecom (Jones & Tsujii which one palticipant in file dialogue asks 1990), and has significantly influcnc(xt the the other participant questions to clarify the research reported here (a similar idea for meaning or intentions of his/her statements) 'automated text composition' in Japanese has tven suggested by Suite & Tomita 1986). (c) user-system dialogues typically initiated by the system, concerning the progress of the D~alogue MT object-level dialogue, disambiguating ambiguous object-level dialogue, i.e. what It is important to emphasize that there is a the user wants to say nexL basic difference between Dialogue Machine Translafion (DMT) 2 systems on the one hand (d) user-system meta-level dialogues typically and conventional MT systems on the olher, initiated by the user, concerning clarification namely the difference of user types. In DMT, of the object-level dialogue, i.e. what was users are dialogue participants who actually ju.~t said. have their respective communicative goals and One of the foreseeable difficulties in DMT is who really know what they want to say. On the how to distinguish these different modes of other hand, the users of conventkmal MT ,are diMogue, that is, how systems can distinguish, typically translators who, though they have first of all, utterances of types (a) and (b) to be enough knowledge about both languages, lack translated and transmitted, from utterances of 'complete understanding' of texts to Ix: type (d) which should not be translated. In translated° particular, dialogues of types (b) and (d) may be This difference in user-types leads to difficult in some cases, because the user posing diffenmt characterizations of interactions questions of clarilication cannot generally betw(~m MT systems and their users. We have recognize whether the difficulties of to mkc into account what this differcnce implies understanding come from 'errors' in translation in designing actual DMT systems. The main or from the other participants' utterances implications can be smnmarized as follows. themselves. For examples of this effect, see Miike et al. (1988). In DMT, the system can ask in thcoc¢ any questions to elicit tile information necessary tot Dialogues of type (c) are found in some form translation which is not explicitly expressed in in most conventional interactive MT systems; the 'source text'. This is impossible in note that with monolingual users such dialogues conventional M-F, because the users do not have are quite different from those found in the 'complete understanding' of the context in 'Translator's Workbench' type of system, since which the texts are prcpmvd, and the users (who it is pmticularly difficult to phrase interactions are translators) simply could not answer such concerning problems of transfer when the user questions. (It is often the case that even human is not expected to know anything about the translalors would like to consult the authors of target language, and when current frameworks the original texts in ordcr to produce a good do not allow us tospecify the relationships translalion.) In order to exploit this advantage in among possible translations defined by different DMT however, we have to overcome several structural correspondence rules. On the other related difficulties. hand, regarding problems with analysis, a particularly useful result of the research on 2 Our concept of DlVlq' should be distinguished from ENtran was to see to what extent potential 'Dialogue-based MT' as proposed by Boitet (1989), in which ambiguities could be recognised on the basis of dialogm; is used to clarify the author's intentions in the structures computed by more or less traditional context of a personal MT system. This is also the case in our DMT, with the crucial difference that the object of parsing techniques (i.e. charts). For dialogues of translation in our case is also part of a dialogue, i.e. the type (c) we are guided by the work of Jones & user's dMogue with a third party. Clearly however, there are Tsujii, mentioned above. sigrfificant areas of overlap between our project and Boitet's. The British Telecom work concerns a system for generating business letters in French,

3 273 German and Spanish on the basis of an appropriate TEE. essentially menu-driven interface (in English). In this scenario the user has taken the The system has a set of preu'anslated fragment initiative in the dialogue, by 'typing in' what pairs some of which have slots for variable s/he wants to say, and having the system find elements to be inserted (e.g. the name of a the appropriate triple. company, or a product) which may or may not Two other scenarios are also possible. In one, be translated in a conventional manner. The the system retains the initiative, and rather like system-user dialogue aims at selecting the in the menu-driven system, selects (or seeks via appropriate target-language expression (TEE) a meta-dialogue) the next appropriate De., and fragment corresponding to some source- then offers a range of appropriate SLES for language expression (SEE) and compiling the selection. In this sense the pair for a TEEs in the appropriate sequence so as to given value of DC can be regarded as a generate the required output. Notice that, since 'conditioned equivalence pair'. the fragments have been pretranslated (presumably by a competent translator), the Finally, in a mixed-initiative scenario, the result is of a guaranteed high quality. user and the system collaborate in the following way: first, a communicative goal is established, This idea is developed in the following ways. and with it a sequence of Des corresponding to First, we assume that the interface menu is the 'dialogue plan'. The user then makes a replaced by a much more complex 'model proposal for the next utterance in the dialogue, dialogue' (see below). In the sense that the and the system searches its database for the pretranslated fragment pairs are associated with nearest apparently appropriate particular points in the model dialogue, they can given the user's input (con'esponding to the SL~) be said to be not just pairs of SEES and TEES but and the DE as given by the dialogue plan. If an in fact triples, since they are identified by a exact match is found, the TEE is generated and description of the dialogue context (DC) which the object-level dialogue continues. However, if conditions the equivalence of the two an exact match is not found, the system gets the expressions, by specifying the point in the user to modify the SLE until it more closely model dialogue at which they are identified, matches the SEE selected by the system. thUS" . It is possible for a given SLE, there may be several TEa depending on the Model dialogue particular IX:, thus: equivalence relation of the two expressions is not guaranteed by the expressions themselves but by the Des which are given rather For example, the English response OK in a independently of the informational content of dialogue may correspond to Japanese the two expressions in the triples. In a context wakarimashita when something is being such as business correspondenc e , it might be the explained, ii desu yo when asserting agreement, case that much less information is necessary to or ijoo desu when it indicates completion of the identify the relevant triple than that conveyed discussion and a change of topic. by the actual linguistic expressions and that, The task of the DMT system can now be because each individual language usually has its divided between first locating the appropriate set own conventions which letters must follow, the of triples involving a given SLE, and then actual informational contents of the two locating the appropriate a't~ for that SEE expressions might be different. The same is true according to the De. of certain types of dialogues. For example, there If we assume that the SLEs are not just are conventional phrases used in Japanese phone 'canned texts', but actually types of text calls (Nagasaki 1971) which, if translated templates of varying linguistic complexity (i.e. literally, would probably mystify the non- from set phrases through to syntactic patterns - Japanese dialogue partner: see below), it can be seen that the first part of Sorry to disturb you when you are busy / the above task can be achieved by traditional eating / about to go to bed / still asleep techniques of parsing or by some other (depending on time of day) matching procedure. The set of different t~cs for a given SLE can be used to trigger a Sorry to have had to disturb you clarification dialogue so as to determine the

274 4 Sorry for having talked too much order capture functional regularities in Excuse me for bothering you communication and to guarantee high qmdity translations. When the system encounters '1'hank you jot going out of" your way to unexpected input, it has a choice of h'yiug to answer the phone steer the user towards input which ix more I assume it is inconvenient for you now, bur.. within its expectations, or to abandon [ am sorry for phoning you without warning temporarily its assurance of high-qtmlity translation ~md operate in a more traditional g wasn't expecting to phone you, but... m~mner. One important research question is what It may be asked why we need the model exactly the oc should look like. Our current dialogues~ file canned text and paratexts, and assumption is that Ix: will actually refer to a conditioned equivalence pairs: would it not point in a 'model dialogue ~, probably a flexible better simply to have a long pre-composition network of script-like structures indicating phase where the writer interacts with an expert possible dialogues that rite system can trmtslate, system which asks lots of questions about perhaps along the lines of work by Wachtel intentions and goals and then uses this 0986) and Reilly (1989). We have not yet knowledge (if require) in a conventional finalised ore" ideas in this area, but we are parse-and-disambiguate system. Of course this considering in particular how to modcl suitably would be ~mother way of addressing the flexible dialogue structures within the domain in problem of under-.specified texts, but it is not question, the problem of intcractions between clear what type of questions could be asked the model dialogues and the recta-dialogues, as unless a speciiic domain of comt×)sition was well as the mechanisms which enable the pin-pointed. This brings us back to domain system to navigate its way through the model knowledge~ which in this case is expressed as dialogue network in response to the user's input. knowledge about what the user can ask next, ~Canned text ~ and extensions which we capture in file model dialogues. It was stated above that the nature of the SLE Conclusion and TEE pairs should be varied. In particular, It is nowadays accepted that we cannot because of the need for tlexibility as compared expect to have fully automatic high-quality MT. to the British Telecom work dcscribcd in Jones We have to dcvclop systems which allow & Tsujii (1990), we assume that tile system will flexible and cffcctivc human intervcntions. Our permit some degree of conventional idca is to explorc diversified approachcs to compositional translation. So SEEs and TLES are interactive MT and in pm'ticular we seek to not always texts, or 'paratexts' (i.e. texts with develop an interactive system for monoling~l slots for proper names or simply translated noun users. Fnrthennore, it seems that several phrases, etc.) but, in some cases, structural interesting ncw approaches become apparcnt descriptions of a more conventional kind. 7his once we escape from the basic assmnption of clearly hnplies that within the system there is a the existence of a concrete source text, and need for analysis (and generation) of the kind explore the idca of 'MT without source texts'. found in conventional MT systems. In particular, where appropriate texts or paratexts are not found for a given input, and the dialogue management part of the system is satisfied that 'free input' is an available option at this point in the model dialogue, then the system becomes more like a conventional MT References system, though with the special characteristics of an MT system which interacts with a H. ARYrA, K. KOGURE, I. NOGAITO, H. MAEDA monolingual user. & H. IIDA (1987) 'Media ni izon suru kaiwa no y6shiki: dcnwakaiwa to kiib6do no kaiwai no For the most part, however, it is assumed that hikaku (Media-dependent conversation manners: there is a stereotyped set of functions involved comparison of telephone and keyboard in performing a global communicative function conversations)'. Jrh6 Short Gakkai 87.34 (Jbh5 in a restricted domain. We can assign surface Short Gakkai Kenkyh Hbkoku, Shizen Gengo representations to these functions which restrict Short 61-NLP-5, 1987.5.22). the form of expression to a certain extent in

5 275 C. BOITET (1989) 'Speech synthesis and Dialo- A. ROTHKEGEL (1986) 'Textverstehen und gue Based Machine Translation'. ATR Symposi- Transfer in der maschinellen Ubersetzung'. In I. um on Basic Research for Telephone Interpreta- B~ttori & H.J. Weber (Hgg) Neue Ansiitze in tion, Kyoto, December 1989. Preprints, 6-5-1-9. Maschineller Ubersetzung: Wissensrepr'asen. H. IIDA (1987) 'Distinctive features of ration und Textbezug, Tiibingen: Max Niemeyer conversations and inter-keyboard interpretation'. Verlag, 197-227. Workshop on Natural Language Dialogue H. SAITO & M. TOMITA (1986) 'On automatic Interpretation, Advanced Telecommunications composition of stereotypic documents in foreign Research Institute (ATR), Osaka, November languages'. Presented at 1st Intemational 1987. Conference on Applications of Artificial R.L. JOHNSON & P. WttlTELOCK (1987) Intelligence to Engineering Problerns, 'Machine translation as an expert task'. In S. Southampton, April 1986. Research Report Nirenburg (ed)Machine translation: theoretical CMU-CS-86-107, Department of Computer and methodological issues, Cambridge: Science, Carnegie-Mellon University. Cambridge University Press, 136-144. M.G. STEER & F.W.M. STENTIFORD (1989) D. JONES & J. TsuJn (1990) 'High quality 'Speech language translation'. In J. Peckham machine-driven text translation'. Third (ed) Recent developments and applications of International Conference on Theoretical and Natural Language Processing, London: Kogan Methodological Issues in Machine Translation Page, 129-140. of Natural Languages, Austin, Texas, June 1990. T. WACHTEL (1986) 'Pragmatic sensitivity in M. KAY (1980) The proper place of men and NL interfaces and the structure of conversation'. machines in language translation. Research I Ith International Conference on Computational Report CSL-80-11. Xerox Palo Alto Research Linguistics, Proceedings of Coling '86, Bonn, Center, Palo Alto, California, October 1980. 35-41. M. KOSAKA, V. TELLER & R. GRISHMAN (1988) HJ. WEBER (1987) Converging approaches in 'A sublanguage approach to Japanese-English Machine Translation: domain knowledge and machine translation'. In D. Maxwell, K. discours [sic] knowledge. Linguistic Agency Schubert & T. Witkam (eds) New directions in University of Duisburg Series B, No.164. machine translation, Dordrecht: Foris, 109-120. P.J. WltlTELOCK, M.M. WOOD, B.J. CHANDLER, J. LEHRBERGER & L. BOURBEAU (1988) N. HOLDEN & H.J. HORSFALL (1986) Machine translation: Linguistic characteristics 'Strategies for interactive machine translation: of MT systems and general methodology of the experience and implications of the UMIST evaluation. Amsterdam: John Benjamins. Japanese project', l lth International Conference on Computational Linguistics, Proceedings of A.K. MELBY (1982) 'Multi-level translation aids Coling '86, Bonn, 329-334. in a distributed system'. In J. Horeck9 (ed) COLING 82: Proceedings of the Ninth M.M. WOOD & B.J. CHANDLER (1988) International Conference on Computational 'Machine translation for monolinguals'. In D. Linguistics, Amsterdam: North-Holland, 215- Vargha (ed) COLtNG Budapest: Proceedings of 220. the 12th International Conference on Computational Linguistics, Budapest: John yon S. MIIKE, K. HASEBE, H. SOMERS & S. AMANO Neumann Society for Computing Sciences, (1988) 'Experiences with an on-line translating 760-763. dialogue system'. 26th Annual Meeting of the Association for Computational Linguistics. Buffalo, NY, June 1988. Proceedings, 155-162. K. NAGASAKi (1971) (Hito ni warawarenai) Kotoba dzukai to hanashi kata. T6ky6: Bunwa Shobo. R. REILLY (1989) 'Communication failure in dialogue: implications for natural language understanding'. In J. Peckham (ed) Recent developments and applications of Natural Language Processing, London: Kogan Page, 244-261.

276 6