COMPUTER-ASSISTED TRANSLATION SYSTEMS: The Standard Design and A Multi-level Design Alan K. Melby Linguistics Department Brigham Young University Provo, Utah 84602 USA ABSTRACT Revision of the raw machine translation by a human translator seems at first to be an attractive The standard design for a computer-assisted way to compensate for whatever errors may occur in translation system consists of data entry of source the raw machine translation. However, revision is text, machine translation, and revision of raw effective only if the raw translation is already machine translation. This paper discusses this nearly acceptable. Brinkmann (Ig8O) concluded standard design and presents an alternative multi- that even if only 20% of the text needs revision, level design consisting of integrated word process- it is better to translate from scratch instead of ing, terminology aids, preprocessing aids and a revising. link to an off-line machine translation system. Advantages of the new design are discussed. The author worked on a system with this stan- dard design for a whole decade (from 1970 to 1980). This design can, of course, work very well. The author's major objection to this ~esign is that it must be almost perfect or it is nearly useless. I THE STANDARD DESIGN FOR A COMPUTER- In other words, the system does not become pro- ASSISTED TRANSLATION SYSTEM. gressively more useful as the output improves from being 50% correct to 60% to 70% to 80% to 90%. The standard design for a computer-assisted Instead, the system is nearly useless as the out- translation system consists of three phases: put improves and passes some threshold of quality. (A) data entry of the source text, (B) machine Then, all of a sudden, the system becomes very translation of the text, and (C) human revision of useful. It would, of course, be preferable to the raw machine translation. Most machine trans- work with a design which allows the system to be- lation projects of the past thirty years have used come progressivelv more useful. this design without questioning its validity, yet it may not be optimal. This section will discuss Here is a summary of objections to the stan- this design and some possible objections to it. dard design: The data entry phase may be trivial if the WHY COMPUTATIONAL LINGUISTS 00 NOT LIKE IT: source text is available in machine-readable form Because even if the algorithms start out "clean", already or can be optically scanned, or it may they must be kludged to make sure that somethino involve considerable overhead if the text must be comes out for every sentence that goes in. entered on a keyboard and proofread. WHY TRANSLATORS DO NOT LIKE IT: Because they feel that they are tools of the sys- The actual machine translation is usually tem instead of artists using a tool. of the whole text. That is, the system is general- ly designed to produce some output for each sen- WHY SPONSORS DO NOT LIKE IT: tence of the source text. Of course, some sen- Because the system has to be worked on for a lonQ tences will not receive a full analysis and so time and be almost perfect before it can be there will be a considerable variation in the determined whether or not any useful result will quality of the output from sentence to sentence. be obtained. Also, there may be several possible translations for a given word within the same gramatical cate- II AN ALTERNATIVE DESIGN gory and subject matter so that the system must choose one of the translations arbitrarily. That There has been for some time a real alter- choice may of course be appropriate or inappro- native to the standard design -- namely, trans- priate. It is well-known that for these and other lator aids. These translator aids have been reasons, a machine translation of a whole text is principally terminology aids of various kinds and usually of rather uneven quality. There is an some use of standard word processing. These aids alternative to translating the whole text -- have been found to be clearly useful. However, na~nely, "selective translation," a notion which they have not attracted the attention of computa- will be discussed further later on. tional linguists because they do not involve any really interesting or challengina linguistic processing. This is not to say that they are I "T& trivial. It is, in fact, quite difficult to per- translation of the segment or an indication of the fect a reliable, user-frlendly word processor or reason for failure of the machine translation a secure, easy to use automated dictionary. But system on that segment. This explains the notion the challenge is more in the area of computer of "selective" machine translation referred to science and engineering than in computational previously. A selective machine translation sys- linguistics. tem does not attempt to translate even segment of text. It contains a formal model of language Until now, there has not been much real which may or may not accept a given segment of integration of work in machine translation and source text. If a given segment fails in analy- translator aids. This paper is a proposal for a sis, transfer, or generation, a reason is given. system design which allows Just such an integra- If no failure occurs, a machine translation of tion. The proposed system consists of two pieces that segment is produced and a problem record is of hardware: (1) a translator work station attached to the segment indicating difflculties (probably a single-user micro-computer) and (2) encountered, such as arbitrary choices made. a "selective" machine translation system (prob- Level 3 provides to the translator all the aids ably running on a mainframe). The translator of levels l & Z. In addition, the translator has work station is a three-level system of aids. the option of specifying a maximum acceptable All three levels look much the same to the trans- problem level. When a segment of source text is later. At each level, the translator works at a displayed, if the machine translation of that seg- keyboard and video display. The display is di- ment has a problem level which is low enough, the vided into two major windows. The bottom window machine translation of that segment will be dis- contains the current segment of translated text. played below the source text instead of the level It is a work area, and nothing goes in it except Z suggestions. The translator can examine the what the translator puts there. The upper window machine translation of a given segment and, if it contains various aids such as dictionary entries is Judged to be good enough by the translator, the translator can pull it down into the bottom segments of source text, or suggested translation~ window with a single keystroke and revise it as To the translator, the difference between needed. Note that writing a selective machine the various levels is simply the nature of the translation system need not mean starting from aids that appear in the upper window; and the scratch. It should be possible to take any exist- translator in all cases produces the translation Ing machine translation system and modify it to a segment at a time in the lower window. Inter- be a selective translation system. Note that the nally, however, the three levels are vastly dif- translator work station can provide valuable feed- ferent. back to the machine translation development team by recording which segments of machine translation Level 1 is the lowest level of aid to the ~re seen by the translator and whether they were translator. At this level, there is no need for used and if so how revised. data ent~ of the source text. The translator can sit down with a source text on paper and begin The standard design for a machine translation translating immediately. The system at this level system and the alternative mul ti-level design includes word processing of the target text, just described use essentially the same components. access to a terminology file, and access to an They both involve data entry of the source text expansion code file to speed up use of connmnly (although the data entry is needed only at levels encountered terms. 2 and 3 in the multi-level design). They both involve machine translation (although the machine Level 2 is an intermediate level at which translation is neededonly at level 3 in the multi- the source text must be available in machine read- level design). And they both involve interaction able form. It can be entered remotely and sup- with a human translator. In the standard design, plied to the translator (e.g. on a diskette) or it this interaction consists of human revision of the can be entered at the translator work station. raw machine translation. In the multi-level de- Level 2 provides all the aids available at level l sign, this interaction consists of human trans- and two additional aids -° (a) preprocessing of lation in which the human uses word processing, the source text to search for unusual or misspel- terminology lookup, and suggested translations led terms, etc., and (b) dynamic processing of from the computer. At one extreme (level l), the the source text as it is translated. The trans- multi-level system involves no machine translation lator sees in the upper window the current segment at all, and the system is little more than an of text to be translated and suggested translations integrated word processor and terminology file. of selected words and phrases found by automati- At the other extreme (level 3), the multi-level cally identifying the words of the current segment system could act much the same as the standard of source text and looking them up in the bilingual design.