QA Before IT

One of the unique aspects of translation products is that there is no one "correct" version; many variations are possible and consequently, almost by definition, clients are usually unable to immediately judge their quality. Perhaps this is why the concept of quality control (QC) has until relatively recently been rather ill-defined, and only loosely applied to the translation business in general.

Consequently, in the early 1980s New Zealand's professional body for translators and interpreters found that indemnity against loss caused by mistranslations was unheard of by insurance companies in New Zealand. Judging translation quality was considered something subjective and controlling the translation process was seen (by translators) as akin to asking novelists to apply a formula to their writing.

There had been a few quality-related rules generally understood by practitioners. One was that translators should translate only into their native language or, if that was not possible, then the resulting translation had to be checked by a native speaker editor. Another was that one should have an appropriate level of subject knowledge in order to undertake technical translation work of a specialized nature. Observance of these rules, however, had been left to the practitioners' discretion rather than to formal enforcement.

In the era before information technology (IT) played a major part in language services, QC was defined as the manual processes carried out by human translators and checkers. These included checking spelling, grammar and figures, as well as translation itself for accuracy and style, etc. In those days, QC was also largely limited to the resources available in-house.

One of the hallmarks of the quality issue of this era was that it was largely dependent on subjective judgments made by translators and checkers, often leaving the client's needs out of the equation. In this sense the language service was removed from the real needs of the customers, and translators often treated their work more like academic exercises; they were in pursuit of "perfect" translations regardless of context and the end use of the product. Such an intuitive and entirely human-based QC procedure was feasible because of the volume of work, the production time, the variety of language pairs, output media and also the final purpose of the translation.

The lack of formal QC measures also had to do with the fact that the translation profession used to be founded almost entirely on "on the job training" rather than on formal institution-based qualifications. In fact, the debate over experience versus formal qualifications is still rampant today, after many specialized translation training schools and academic institutions have been established the world over.

Because of the lack of official "measurements" of their skill levels and their outputs, and despite the specialized nature and the expertise required, the work of translators had been cast somewhat outside the norm of so-called "professionals" such as medical doctors, lawyers, engineers, etc. This situation has not been helpful for either translators or customers; the former were often undercut by amateurs or had to face clients who expected unrealistically low rates while the latter had to persevere with jobs of less than satisfactory quality or suffer the consequence of poor translations. Under these circumstances, quality assurance (QA) in any formal sense in the translation industry was almost non-existent.

By comparison, from its birth in the 1950s, machine translation (MT) has worked on a totally formalized basis with language analysis and generation rules. Even inputs need to be regulated (by way of pre-editing of input text) for better results. In this sense MT can be relied upon to invariably apply whatever rules are programmed into its software and to output consistent terminology, etc. This environment would seem to imply that it would be easier to apply QC procedures to MT productions than to those done by humans alone.

And yet the very inflexibility in changing "rules" is often responsible for poor quality outputs. Human translation production processes have been more or less reliant on the individual style of each translator and the mere thought of applying any standardization would have been scoffed at by many practitioners. This thinking may be reflected in the long-held attitude of many translators towards MT: computers cannot undertake the translation process, because it is only possible by means of human intelligence. In other words, this human mental process cannot be formalized.

At the same time, the end users of translations had almost the opposite and unrealistically optimistic expectations that MT would resolve the world's language problems once and for all. To the delight of many human translators, however, the early results of MT only supported the view that the task was beyond the computer's capabilities.

QA After IT

In the mid 1980s when IT started to impact on the translation business with multilingual word processing, DTP and faxes, for example, the QC process began to evolve from an entirely manual and intuitive style to one following systematic procedures and the use of technology. While the computer began to be used for general job management purposes, specific IT tools were integrated into QC procedures. Electronic spelling checkers were used as part of standard document preparation (although not replacing a human proofreader entirely) and the fax made possible the concept of remote "in-country" translation and checking. This also meant that some clients were able to assess the quality of translation with the help of their in-country contacts.

For the suppliers, the new IT applications meant a new capacity to be able to add value to a straight translation job by the use of DTP, for example. This, on the other hand, created a new area to be covered by QC. In addition to checking translated work in the traditional sense, extraneous factors such as formatting, fonts, graphics, etc. also needed attention. In this way, QC had to be extended to the whole operation of translation production rather than just the translation process.

With the advancement of IT and the increasing globalization of world markets together with enhanced translation capa-city, translation work started to increase both in volume and in the variety of work content, often with a reduced production time. Under these circumstances the need for project management and formalized QC procedures became essential.

Over this period translation services also became much more sensitive to customers' real needs, and the idea of "quality" was no longer always taken in a purist sense but in the context of customer requirements. For example, draft translations with a quicker turnaround at a reduced price were wanted by some customers for certain jobs. This supplier awareness in turn probably stimulated more demand for translation (which would otherwise have gone unrealized). In fact, this "information-only" translation need is one of the markets specifically targeted by some MT developments.

Into the early 1990s low-price desktop MT software (sometimes called PCMT) emerged in the market, while online MT services were also in operation. Some translation operators used MT systems in an attempt to increase productivity. By this time it had become clear to both translation suppliers and their end users that while MT could not be expected to produce perfect translation, perhaps it would have its place. End users might take advantage of speed and cost factors for information-only purposes or for first-pass translations before deciding on the need for more precise work by human translators. For the translation suppliers, some repetitive text and large-volume jobs in certain technical domains with a short production deadline became possible candidates for MT processing. Translator's workstations and CAT (computer-assisted translation) were also being used in some translation offices.

This was also the time when the localization business established itself in a major way in the translation sector. This new field of translation represented the coordinated skills of compu-ting and translating and in many ways started to bring in a more IT-oriented approach.

In the mid 1990s the impact of the Internet was felt strongly by the translation industry, bringing a new generation of "teletranslation", whereby customers and service suppliers are linked electronically on a global scale. Today there are a large number of Web-based teletranslation services in operation, including both MT-based and human-based services. With the former the user knows the trade-off between the cost/time and the quality factors, and this understanding makes the transaction mostly straightforward in terms of QC for translations; the service comes with a "disclaimer" on the quality of the translation.

While the first introduction to teleworking for translators began with fax, they needed to wait for today's more mature telecommunications environment to provide sophisticated worldwide "virtual" services. The virtual service formula seems to be largely operable with translation work; text arrives and departs between the client's and the supplier's screens. The advantage of teletranslation is that the customer can tap into a translator who may be best qualified to take on the given translation assignment without being confined by locality.

Physical distance in human-based teletranslation services (which cater to wider translation needs than their MT counterparts) does, however, have implications for the quality issue. For example, electronic links inadvertently break down from time to time, sometimes without either party realizing. Also, even when it is working, the question remains: does the electronic link allow a sufficient feedback channel between the customer and the supplier or the supplier and his/her subcontractors?

Translators often need to ask questions regarding the text after accepting a job, in order to resolve the meaning of ambiguous sentences, check the spelling of proper names, clarify jargon, etc. Clients may not be able to get on the phone to demand a delayed job to be dispatched immediately. Suppliers will require a new QC system when a large volume of translation work is organized in a distributed manner. QA for teletranslation needs to take into consideration these extra factors due to the very fact of "facelessness" in their style of providing services.

The Future of QA

Having looked at some aspects of how quality issues evolved in the translation industry, perhaps it's time to try to define what we mean by quality in the context of translation service, and then to explore future strategies regarding the quality issue in general.

The five categories of quality by Garvin quoted in Samuelsson-Brown [1] and given on the next page provide a basis for discussion:

1. Perceived: this is the 'you know it when you see it' view.
2. Product-based: this is based on quality being design-oriented, precise and measurable.
3. User-based: this is based on the 'fitness for purpose' from the client's perspective.
4. Operations-based: this is based on the principle of 'conformance to specification' i.e. error-free.
5. Value-based: i.e. 'best value for money for a given purpose'.

QC in the translation business used to be driven mainly by the subjective judgment of translators and their checkers. In this sense it was closest to the "perception-based" approach, and depended mainly on one's linguistic skills. This remains true today as far as translators' self-checking is concerned. However, with IT applications and with the translation sector taking on the characteristics of a service industry, the overall QC strategy by practitioners has now moved to a less intuitive and more systematic approach. Of all translation businesses, the localization sector may be considered the forerunner that first attempted a formalized process in QA, based on all three factors of "product", "user" and "operations". This head start may be partly due to the nature of localization work, but is also perhaps due to the "culture" of a systematic approach inherited from the computing industry background combined with relatively obvious and fast feedback from the user market.

In the past, many attempts to quantify "translation quality" have been made by the world's translation associations as well as by individual translators, mainly with the objective of improving the professional status of translators—this had little clear, tangible success. Today, however, commercial pressure more than anything else seems to be pushing this. For example, the applicability of ISO 9000 standards has been a subject of debate among practitioners in recent years.

In the time to come the translation business will increasingly need to apply this kind of formalized process, treating translation as a "product" or "commodity". Customers will be increasingly looking for QA, and this will be even more true with teletranslation services in order to compensate for their "virtual" nature. How will QA evolve and what factors will improve the current model? The answer will most likely lie in overcoming the bottleneck of the human process (the translation process in particular) in a QA model.

One of the beliefs held by QC guru Dr. Deming was simply to get it right first time. This motto was hammered into those of us working in a New Zealand government translation bureau in the early 1980s. We were told that the translator should always aspire to get the right translation first time around rather than leaving the effort to find a correct translation till later. This advice is sometimes difficult to follow, particularly with the flexibility afforded by word processing which can encourage translators to put "something" down even if they do not understand the word or phrase correctly in their first attempt at a draft translation.

Recently I was interested to hear that a local translation firm decided to use more experienced translators to do the first draft and the less experienced to edit/check their work. This sounds paradoxical to some practitioners, but it in fact follows the Deming logic. A senior translator is less likely to make errors than a junior translator; thus, by getting it right first time the margin of error will be minimal. Doing it the other way around takes more time for editing and checking and possibly produces lesser quality work.

Looking into the future, it is inevitable that MT will progressively encroach upon the human translation process. How effectively can MT be implemented in relation to QC and QA issues? Applying the Deming formula, perhaps an MT system can be developed and specifically designed to edit or check a good-quality base translation. For example, a senior translator does a first draft translation. The text is then put through MT which back-translates the text into the source language. The result is then compared with the original text for numbers, proper names, number of sentences (paragraphs), terminology, grammatical structures, etc. Further research is necessary to investigate how commercially feasible this scenario may be.

One of the world's MT authorities, Professor Nagao of Kyoto University, once said that both birds and planes can fly and yet they fly based on totally different principles. MT does not necessarily follow the way humans translate. In fact, MT's advantage lies in its very mechanical and systematic methods (think of its processing speed, capacity and consistency in the use of terminology). The human translation process will remain extremely difficult to subject to a total QC procedure and yet MT will readily lend itself to such a formalization.

While considering how to improve QA for translation services, we have to assume that human factors will always remain opaque. The challenge is then to find a way to minimize the opaque factor and counter it with the machine's transparency. For example, the use of CAT or translation memory clearly assists "systematizing" the human translation process, making it easier for benchmarking and achieving consistent quality. The extreme example would be fully automated on-line MT services, which can be subjected to a QC system. Then the question is to what extent one can afford to remove the "human components".

During my recent participation in PACLING '97 (Pacific Association for Computational Linguistics, held in Tokyo) listening to computer and AI (artificial intelligence) scientists present their research into natural language processing technology, I realized that the more computers start behaving like humans the more their translation quality may improve. At the same time, they will make more human-like mistakes. In this sense we will never have "perfect translation" either by machines or humans.

There will be no such thing as "perfect QA" on the human side either, but by the very effort of scrutinizing the translation production process for QA purposes we will come closer to finding a workable model. Comments given by translation practitioners in relation to adopting ISO 9000 are that its applications may sound too demanding to be realistic, but the process of preparing for the certification does help clarify inefficiency and areas for improvement [2] [3]. In the long run this may be a worthwhile exercise. Last but not least, a successful formula for QA requires input from clients. Striving for higher quality needs to be a joint effort.

References

[1] Garvin, D. 1988, Managing Quality, Free Press. quoted in Samuelsson-Brown, G. 1996, ISO 9000 - Moral Blackmail or a Useful Tool? in XIV World Congress of the Federation Internationale des Traducteurs Proceedings, vol. 1.

[2] Samuelsson-Brown, G. 1996, Ibid.

[3] Muzii, L. 1997, Italians Clash Over Translation Standards, in Language International, 9.3.

---

Minako O’Hagan
School of Communications and Information Management,
Victoria University of Wellington,
New Zealand
E-mail: Minako.OHagan@vuw.ac.nz