Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
The Commercial Challenges CURRENT STATUS OF MACHINE TRANSLATION DEVELOPMENT How, where, and by whom are machine translation technologies and machine aids for translation being developed? Who is using them and for what purposes? Before examining Japanese to English machine translation as it is being developed and used in the United States and Japan, it is helpful to review the broader context of machine translation development throughout the world. Machine translation development has advanced rapidly in the past few years, and while numerous challenges remain, machine-assisted translation is no longer a dream, but is actively and increasingly used around the world. Companies and governments are developing and/or using machine translation technologies in the Americas, Europe, Japan, and in the newly industrializing countries of Asia. Machine translation and machine-aided translation are now being used in organizations as diverse as translation bureaus, multinational corporations, and government defense departments. These technologies help to scan information, collect intelligence, translate product manuals for export, and improve the efficiency of professional translators. The overview below provides a basis for understanding the organizational, political, financial, and societal reasons for differences in their development and use in the United States and Japan. The major systems in use today are not, however, aiding translation from Japanese to English; most are performing what is considered by many the easier task of translating from one Western language to another. Many examples can be found in the public or government sector. The Pan American Health 8
9 Organization, for example, uses its internally developed system to translate health and agricultural information between English and Spanish. The U.S. government, including the Air Force's Foreign Technology Division, has long used Systran to translate Russian, German, and French into English for intelligence purposes. The government is currently supporting Systran's development of a Japanese language system. Similarly, the Canadian government uses the Logos system for translating English to French at the Departments of National Defense and State, among others, and uses METEO to translate weather forecasts. The Smart Translator is used to announce job openings. The European Community is supporting a large initiative aimed at developing a system to translate among all the official languages of the Community. In the private sector, manufacturing companies are using machine translation technology to produce product manuals and translation companies are using it to improve the volume and efficiency of translation efforts. Xerox, for example, uses the Systran machine translation system to translate its photocopier product manuals for distribution throughout the world. A European computer maker, Nixdorf, uses the Logos system to achieve the same end. Translation companies such as Lexitech in Canada also use Logos, and ALPNET has developed a worldwide network of translation services based on its own internally developed interactive and machine-aided software. METAL has a number of clients in Europe and ATAMIRI is used by Wang for the translation of product manuals into several languages. There are a few examples of machine translation technology used for Asian languages. Logos, one of the oldest U.S. machine translation companies, got its start in 1971 when it developed an English to Vietnamese system for the U.S. Defense Department. Currently, the Defense Department is funding machine translation from and into Korean. It can be seen from the examples above that the world of machine translation spans the globe and includes efforts in many language pairs for a variety of uses. Some of the oldest, most successful, and well-established machine translation companies are American or are based on technology developed in the United States. Systran is perhaps the grandfather of U.S. machine translation companies and its efforts in language pairs other than those involving Japanese continue to be U.S.-based. The next oldest commercial machine translation system is that of Logos, also an American company. One of the best-known Japanese machine translation companies, Bravice International, which claims to have sold 4,500 software units, bought its technology in the United States through the acquisition of a U.S. firm. It is useful to consider this context as one examines the more specific case of Japanese to English machine translation. Who, then, are the developers of machine or machine-aided translation between Japanese and English? The major commercial efforts in the United States are being conducted by a handful of companies. As mentioned above,
10 Systran, with U.S. government support, is supporting the Japanese-English combination; it should be noted, however, that Systran's commercial Japanese efforts have been Japanese-ownel Some large U.S. corporations, including IBM and DEC, are pursuing internal research in this field, but their Japanese to English efforts are located in Japan.7 IBM's English to Japanese machine translation system, developed for internal use in the corporation's Tokyo facilities, is currently in operational use for the translation of computer manuals. There are some small-scale efforts, experiments in assembling and using machine translation technologies for very limited domains. Examples are Smart Communications and a small ongoing effort at the Veterans Administration Medical Center in Baltimore that uses public domain software to translate a narrow range of medical documents between Japanese and English. Efforts in machine-aided translation development are currently underway at ALPNET and LinguaTech, both of which are based in Utah and originated at Brigham Young University. Logos, which continues to invest heavily in machine translation development, has conducted research in Japanese but has never undertaken development of a Japanese system. There is also considerable theoretical research under way among computational linguists at U.S. universities, such as work at Carnegie Mellon University's Center for Machine Translation, and the Linguistics Research Center at the University of Texas, and the sub-language approach pursued cooperatively by Hunter, Monmouth, and N.Y.U. If basic research on natural language processing and computational linguistics is taken into account, the United States still maintains a signficant research effort.8 What is lacking in the United States is a strong development effort on Japanese to English machine translation. The world of Japanese to English machine translation development in Japan offers a sharp contrast to that in the United States. Every major Japanese computer or electronics firm has invested considerable effort in machine translation research and development and many claim to have introduced workable systems. Without evaluating quality, it is nonetheless significant that there are at least twenty operaiional9 systems in Japan that translate from English to Japanese. Operational systems in Japan that translate Tom Japanese to English 7 At its Tokyo Research Lab, IBM is working on a Japanese to English machine translation system for use in translating newspaper editorials and economic materials, which is now in a research prototype stage. 8 It should be noted, however, that research in computational linguistics that is not related to machine translation will not necessarily contribute to machine translatic n. On the other hand, the ease can be made that work on machine translation serves as a test bed and stimulus for Adler kinds of natural language processing investigation. 9 There is considerable ambiguity about what constitutes an "operational" or "usable" system. The sheer number of Japanese developers who even claim to have an operational system is nonetheless an indication of the comparatively strong interest and resources devoted to machine translation in Japan.
11 are available from NEC, Fujitsu, Oki Electric, Bravice International, Sharp, Toshiba, Hitachi, and Sanyo Electric to mention a few of the companies. Japanese to English systems are also under development at NTT, Mitsubishi Electric, KDD, and Toshiba. In addition, the Japanese government has supported an important effort in machine translation development. This effort, which involves the Ministry of International Trade and Industry's (MITI) Electrotechnical Lab, the Science and Technology Agency's (STA) Japan Information Center of Science and Technology JICST), and Kyoto University, was started in 1982. The Electronic Dictionary Research Project conducted by MITI in connection with the Fifth Generation Computer Project aims at the development of a detailed dictionary with more than 200,000 words and multiple usages.~° Supported by the Ministry of Post and Telecommunications, research is underway at ATR on speech translation telephony. Technology transfer to industry has been made possible throughout these projects via industrial participation. The projects feed into the effort at JICST for translating science and technology abstracts. These projects are developmental vehicles that spin off nationwide results; their continuity and commercial emphasis build capability in the companies. Who, then, is using Japanese to English machine translation? If the typical pattern is for translation to be done in the country of the target language, then Japanese to English machine translation should be done most efficiently in the United States, rather than Japan. At the present time, however, this is not the case. Most Japanese to English machine translation is conducted in Japan. The fact that many Japanese computer makers have developed their own machine translation capability reflects their orientation toward product exports and their need to control the quality of translated manuals. It also reflects an understanding of the importance of machine translation technology and its possible spin-offs to the information industry as a whole. Japanese-developed Japanese to English systems are not widely used in the United States. Lack of compatibility between hardware and software is one of the impediments. Microelectronics and Computer Technology Corporation (MCC), a private U.S. microelectronics and computer science consortium, is a relatively new user of Japanese to English machine translation (MCC is using a Japanese-developed system) and has begun to use the technology in a particularly forward looking manner that will be discussed in more detail below. The University of Wisconsin's Biotechnology Center planned to use a Bravice system 10 The 10-year EDR project that began three years ago includes an English dictionary, a Japanese dictionary, and a neutral dictionary that connects the Japanese and English dictionaries. 11 See Hitoshi Iida, "Advanced Dialogue Translation Techniques," ATR Interpreting Telephony Research Laboratories, ATR Symposium on Basic Research for Telephone Interpretation, December 11-12, 1989.
12 on an experimental basis to scan and translate Japanese language databases on biotechnology, but has been unable to do so because of the high cost of the necessary hardware. Currently, there is no U.S.-developed Japanese to English machine translation system on the market. The U.S. defense and intelligence community will, in all likelihood, be the first major user of Systran's Japanese to English system when it is fully operational. MARKET PROSPECTS Today, the volume of machine-translated documents remains comparatively small. In contrast, some Japanese believe that the annual market for all translated materials in 1988 was about 800 million yen, and that the quantity of translation will double over the 1990-1992 period. One developer even estimates that by the year 2000 there will be 500,000 to 2 million machine translation systems in use throughout the world,~3 assuming that substantial improvements are made in intermediate systems that can run on small personal computers. According to this vision, international businessmen will need small computers with built-in machine translation systems. At the same time, Japanese experts note that large Japanese companies working on machine translation do not believe that this business will yield great profits, at least in the short term. They do, however, see machine translation as a mechanism for learning more about natural language processing technology in general, which they judge to be a key technology in the next century. Despite the interest in machine translation technology, profits in the United States and Europe are very slim, if there are any at all. While there are companies developing systems for internal use, the independent developers are (as noted above) few in number. This contrast with the situation in Japan may be explained in a number of ways. Critics of machine translation argue that the products are based on dated technology. They argue that translators harbor serious doubts about machine translation on quality grounds, even if translators have a hard time quantifying the concept of quality. Translation, particularly in Europe, is poorly integrated with office and publishing computer environments where the potential benefits of machine translation could be substantial. Prom a Japanese viewpoint, however, it is possible to create demand. Viewed from this angle, the more machine translation systems are made available and put into practical use, the more the demand will grow. Japanese industry and government are willing to plow large investments into Japanese to English ]2 See Japan Electronic Industry Association, A Japanese View of Machine Transition. . ., op. cit., p. 5 and Appendix 5. 13 Many experts see this as a very high estimate.
13 machine translation. It is estimated that a four- to five-year effort involving 50 to 70 people is needed to develop a general-purpose mainframe system for delivery, at a cost of $13 million for the entire period.~4 This is, however, only the beginning of the investment that is necessary. After delivery of the system, considerable resources must be invested in maintaining and improving it in response to user complaints and needs. In short, the required investments are so large that most companies find it impossible to recover costs by selling only hundreds of systems. This is a primary reason for Japanese government support not only of researchers, but also of commercial developers, although there are other important reasons such as reducing the language barrier between Japan and other countries and disseminating Japanese technical information worldwide. Given that the major commercial machine translation systems have been mainframe-based, the traditional markets for machine translation have been limited to translation bureaus, multinational corporations, and intelligence and information gatherers, particularly in government. Developers have identified two major targets for development. One is the large-scale, general-purpose, mainframe-based system for use by big companies and governments. The goal of this type of machine translation system using large-scale hardware is high volume translation of documents, sometimes for mass distribution. Even those involved in development work on general-purpose, publication quality machine translation systems say that such high-quality, general-purpose systems are 10 to 20 years over the horizon. The second is a small-scale machine translation system for use by small to medium size companies and even individual researchers. The purpose of this type of smaller scale system is to translate for specialized applications. Many Japanese commercial developers are focusing on small-scale system development. There is a third kind of development that is intermediate between We two systems mentioned above.~5 From a commercial perspective, it is this kind of intermediate system development that is seen by the Japanese as the only feasible target for at least the next decade. Such systems are in use today for information scanning in resmcted domains. In developing and enhancing machine ~anslabon systems, developers stand to benefit from close interaction with users. Japanese developers rarely consult with users on the details of systems design, but they do seek out users' views on what features are needed in pre- and post-editing and on general issues of man- machine interface. Japanese developers prefer to interact intensively win a limited number of users so that they can respond effectively to their needs. A 14 Makoto Nagao, written response to "stimulation questions" prepared for the Symposium on Japanese to English Machine Translation. 15 Systems like those at Xerox run on microcomputers and are specialized, but cannot be considered "small-scale" since they translate more than 10 million words annually.
14 conscious strategy of selling a limited number of systems is often pursued in order to facilitate this process. New users benefit later from this accumulated feedback embodied in system configurations developed and perfected for other users. Although Japanese observers complain that mechanisms for exchange of information between developers and users are inadequate, in comparison to the situation in the United States there has been closer interaction among the research, development, and user communities in Japan. One of the limiting factors on the machine translation market is the fact that most Japanese systems operate on only one type of hardware. In fact, some large mainframe computer makers pursue machine translation development as a strategy for marketing their hardware to large companies. Bravice produces the only Japanese to English machine translation system that can be used on a variety of hardware. It runs on most small microcomputers sold in Japan. A practical barrier to widespread usage of small-scale Japanese-developed machine translation systems is the limited availability of hardware that supports Japanese characters (kanji, hiragana, katakana) outside of Japan. International competition in machine translation is not mature. Given the barriers to hardware interoperability, the most prominent examples of competition occur among large-scale communication systems. In 1988, a number of computer companies competed for contracts associated with the development of a large communication system for the Korean and U.S. armies and some machine translation system developers received contracts. Reflecting the practical limitations on machine translation technology today, there are a number of efforts that focus on combining machine translation technology and human translation. ALPNET has developed a strategy based on the assessment that machine translation is not the solution to all user needs, and that it is not an all-or-nothing alternative. Technology tools now in existence (such as multilingual word processors, dictionary look-up systems, character recognition and word processing systems) can be effectively used by human translators at much less cost than fully automated machine translation systems. New ''linguistic engine" options can be added to a user's existing applications, thereby increasing ease of use without forcing the customer to learn new user interfaces to complex systems. The experiences of developers like ALPNET illustrate the fact that successful application of machine translation and related technologies depends on an understanding of user needs, establishment of expectation levels, matching the tools to the job, and the training of skilled professional people to use the tools. Another possible approach is to redesign systems using new technology and locate smaller, modular systems in places such as schools where translators are 16 Exceptions to this general statement about interactions in the United States include relationships between developers and users such as Systran and the U.S. Air Force or between Logos and AT&T.
15 being trained. Also, a broader appreciation of machine translation and related technologies to cover a wide range of communications problems could enhance the use of currently existing tools in solving practical problems. For example, many potential users in Europe are not professional translators but secretaries working for large multinational corporations who must write correspondence in foreign languages. Market prospects could be broadened by integrating linguistic tools into the office environment, the publishing business, and even He engineering development environment. If we consider particular applications of machine-aided translation technologies to specific consumer products, demand may indeed grow significantly over the next few years. A composition aid for technical writers, for example, could improve understanding of cultural nuances, not to mention grammar and syntax. Another product application that illustrates the point is a hand-held electronic phrase book for use by foreign travelers. Such applications, it should be noted, offer promise, but they remain distant from the machine translation heaven of general-purpose, high-quality systems. USER NEEDS People who want to read translated material care little whether the translation is done by people or machines or a combination. They simply want reliable translation that is cheap and fast. For some of these users, such as those who must scan very large volumes of information in order to follow trends in research and development or identify documents for full translation, the speed of translation is important. Turnaround time is, in the experience of the users of the Air Force's system, more important than quality, defined in terms of naturalness of expression and precision in conveying meaning.~7 Other large volume users, such as companies that require translation of their operating manuals, require precise and understandable translation- but in fairly limited domains. The classic example of such a user in the United States is Xerox. Xerox has been using machine translation for more than a decade to translate technical documents. These are large volume projects which involve high reproduction and updating rates. For large companies that sell a wide range of products in global markets, translation costs represent a significant component of new product expenditures. The Xerox experience deserves further mention. Xerox has developed an approach that integrates desktop editing programs and a Systran machine translation system to produce service manuals, Gaining programs, and operator 17 Some argue that quality standards must include a mixture of turnaround time, accuracy, and readablity, the relative importance of each varying with the needs of the particular user.
16 manuals for use in Europe, Latin America, and North America. This approach, which has evolved over time, involves the training of technical writers and pre- editors in a set of simple writing tools (Multinational Customized English) developed by Xerox. Dictionary building is constantly under way as technical writers and translators from different parts of the world, joined through the Xerox worldwide network, submit new words and phrases. At Xerox, machine translation has shown concrete results. Significant improvements have been seen in productivity per finished page of translated text. Producing more than 40,000 pages annually of documentation translated through this process, Xerox now finds that translation is no longer a barrier to product launch. Machine translation enhances the company's ability to introduce new products almost simultaneously all over the world. Xerox found that management initially had high expectations for what machine translation could do, while linguists were skeptical. Over time and with growing experience, the perceptions of these two groups have begun to converge. (See Figure 2.) By gradually developing a system that translators are able to see as a tool in their work, and one that produces high-volume product in restricted technical domains, machine translation has been fully integrated with business operations. In contrast to large users, there are countless potential users of small machine translation systems who have different kinds of needs. In many instances, these are researchers (scientists and technical personnel) in the United States who need to know about developments in Japanese science and technology. For many of them, readability and smoothness of translation is less important than timeliness and access to a narrowly targeted range of technical documents of interest. High Expectation Level Low - / Managers - - - - - Linguists 1 1 To l me · Reality FIGURE 2 Expectations versus reality of computer translation. SOURCE: Mana Russo.
17 MCC is now experimenting with machine translation to monitor overseas technology developments and provide translated materials to member firms. For an investment of $50,000, MCC has assembled a machine translation system that includes an optical character reader, a Japanese-made personal computer that is also used as a JICST terminal, a Japanese to English machine translation software package, and a U.S.-developed workstation. MCC encountered problems in connecting the computer systems developed in Japan and the United States and in getting adequate vendor support for the machine translation system developed in Japan. Fifty percent of the sentences in the trial output have been judged to be accurate. For a modest investment ($50,000 to $60,000), the machine translation experiment at this U.S. industry consortium is considered to be worth the effort. MCC's International Liaison Office expects to augment its already successfully established functions of monitoring developments in Japan and providing technical support to its researchers by using machine translation first to translate titles and, eventually, abstracts and short papers. MCC has, in addition, recently initiated a research effort on knowledge-based natural language processing targeted toward machine translation technology development. Beyond corporations, individual professional translators and researchers also qualify as potential users of machine translation systems and machine aids for translators. Companies such as LinguaTech are developing workstations for translators who can work at home using a microcomputer with a high-capacity disk drive, a printer, and a port for telecommunications. A translator can receive source texts and glossaries with specific subject matter along telephone lines. A variety of options are available to the translator, including composition tools, a bilingual corpus that permits text retrievals when necessary, multilevel terminology files, and optional use of machine translated texts. To use a baking analogy, machine translation is one ingredient in a range of elements that the translator (baker) can combine in myriad ways to accomplish his work. The baker may wish to use the heavy-duty bread mixer (machine translated text) for some tasks, but he also keeps his drawer of spoons (manual tools).~9 Potential users, particularly individuals or small companies, may find it hard to learn about machine translation. In Japan, potential users learn about machine translation through newspapers and television as well as through demonstrations at computer company service centers. Only limited information is obtained Trough these channels. The potential Japanese user is, however, in a much better 18 In judging the value of machine translation, it must be remembered that, unlike the case with European languages with cognates and familiar script, Japanese is totally unintelligible to the potential English-speaking user. Some, therefore, believe that capturing even 50% of the Japanese text can be a useful step forward in communication. 19 This analogy is developed by Alan K. Melby in his article "Recipe for a Translator Work Station," Mull ilingua, 3-4 (1984), pp. 225-228.
18 position than his U.S. counterpart. For the U.S. translation community, demonstrations and meetings of professional organizations provide channels for transmitting information that augments the articles appearing in specialist journals. For other users who have no direct contact with the translation community (such as researchers and people in small business), acquiring information about machine translation and making judgments about purchase and use are even more difficult. We shall return to this point again in the final section on R&D policy. JAPANESE AND U.S. USERS: CONTRASTING NEEDS As we think about what type of Japanese to English machine translation system would help users in the United States, it is important to keep in mind that their needs are different from those of most users in Japan. The profile of the typical user of Japanese to English machine translation in Japan is a large corporation engaged in global exports. This user can "control" the source document because it is in his native language and pre-editing is easier in this case. This user can customize the system to suit his or her needs and tolerate marginal machine translation because there is a direct cost justification for this effort. The profile of the typical user of Japanese to English machine translation in the United States is someone who needs expanded access to Japanese-language technical information but who has no fluency in Japanese. This user has no control over the source text and little ability to pre-edit. For monolingual users, the "raw" output must be more reliable and accurate than for the typical Japanese user. For the user in the United States, there are many problems associated with gaining knowledge of and access to Japanese databases, inputting text, and post- editing requirements. In view of the broad-based needs of the U.S. user, marginal machine translation is in many cases not useful. (See Figure 3.) However, there are many possible uses of machine translated text now available, particularly if some post-editing is done, for scanning information. Japanese to English machine translation systems developed in Japan do not, in the opinion of some leading U.S. developers, fit the needs of many potential high-volume U.S. users. The requirements are for high-quality, broad-domain systems which will be based on new technology. What are our choices? Should we wait for Japan to develop high-quality Japanese to English machine translation systems in the year 2000 and later? After all, Japan has a head start and Japanese is their language not ours. Or should we develop Japanese to English machine translation systems here in the United States? Are there other alternatives? Answering these questions requires an understanding of the technical challenges facing those engaged in R&D on machine translation.
19 High~C)uality Systems Cost-Effective Systems Prototype Systems Demonstration Systems JAPAN: Exporter's USA: Researcher's production tool information utility ~, · doesn't control source · controls source document _ document · can customize system to _ · needs read-only narrow domain broad-domain system tolerates high-cost, · needs low-cost, usable raw low-quality, raw output output ~_ ~_ (I) · Special Purpose Systems · Interactive Systems Domain- · General Specific Purpose Constrained Systems Systems Systems Japanese to English machine translation systems useful for users in Japan only Japanese to English machine translation systems useful for users in both Japan and the United States \ Japanese to English machine translation systems useful for J users in Japan and possibly the United States FIGURE 3 User profiles for Japanese-English machine translation. SOURCE: Bernard Scott, Logos Co~oranon.
