Chapter 5

In 1966, the U.S. government-issued ALPAC (Automatic Language Processing Advisory Committee) report offered a prematurely negative assessment of the value and prospects of practical machine translation systems, effectively putting an end to funding and experimentation in the field for the next decade. It was not until the late 1970s, with the growth of computing and language technology, that serious efforts began once again. This period of renewed interest also saw the development of the Transfer model of machine translation and the emergence of the first commercial MT systems. While commercial ventures such as SYSTRAN and METAL began to demonstrate the viability, utility and demand for machine translation, these mainframe-bound systems also illustrated many of the problems in bringing MT products and services to market. High development cost, labor-intensive lexicography and linguistic implementation, slow progress in developing new language pairs, inaccessibility to the average user, and inability to scale easily to new platforms are all characteristics of these second- generation systems."

As explained in August 1998 by Eduard Hovy, head of the Natural Language Group at USC/ISI (University of Southern California/Information Sciences Institute), machine translation implies "language-related applications/functionalities that are not translation, such as information retrieval (IR) and automated text summarization (SUM). You would not be able to find anything on the Web without IR! -- all the search engines (AltaVista, Yahoo!, etc.) are built upon IR technology. Similarly, though much newer, it is likely that many people will soon be using automated summarizers to condense (or at least, to extract the major contents of) single (long) documents or lots of (any length) ones together."

= Experiences

In December 1997, AltaVista, a leading search engine, was the first to launch a free translation software with Babel Fish -- also called AltaVista Translation --, which could translate webpages (up to three pages at the same time) from English into French, German, Italian, Portuguese or Spanish, and vice versa. The software was developed by SYSTRAN (an acronym for System Translation), a company specializing in machine translation software. SYSTRAN's headquarters are located in Soisy-sous-Montmorency, near Paris, France. Sales, marketing, and research and development are based in its subsidiary in La Jolla, California.

This initiative was followed by other translation software developed by Alis Technologies, Globalink, Lernout & Hauspie, and Softissimo, with free and/or paid versions on the web.

Based in Montreal, Quebec, Alis Technologies has specialized in development and marketing of language handling solutions and services, particularly language implementation in the information technology industry. Alis Translation Solutions (ATS) has offered applications in a number of languages, and tools and services to improve the quality of translations. Language Technology Solutions (LTS) has marketed advanced tools and services for language engineering and information technology (90 languages covered).

Based in Ieper, Belgium, and Burlington, Massachusetts, Lernout & Hauspie (L&H) was a leader in advanced speech technology for commercial applications and products, with four core technologies: automatic speech recognition (ASR), text-to-speech (TTS), text-to-text (TTT), and digital speech compression (DSC). Its ASR, TTS and DSC technologies were licensed to companies in telecommunications, computers and multimedia, consumer electronics and automotive electronics. Its TTT translation services were provided to IT companies, and vertical and automation markets. The Machine Translation Group created by Lernout & Hauspie included L&H Language Technology, AppTek, AILogic, NeocorTech, and Globalink. Lernout & Hauspie was later bought by Nuance Communications.

Globalink, a company created in 1990 in the U.S., focused on language translation software and services, i.e. customized translation solutions built around software products, online options, and professional translation services. The software products were available in Spanish, French, Portuguese, German, Italian and English, for individuals, small businesses, multinational corporations and governments, from a stand-alone product giving a fast draft translation to a full system managing professional translations.

As explained on the company website in 1998, "with Globalink's translation applications, the computer uses three sets of data: the input text, the translation program and permanent knowledge sources (containing a dictionary of words and phrases of the source language), and information about the concepts evoked by the dictionary and rules for sentence development. These rules are in the form of linguistic rules for syntax and grammar, and some are algorithms governing verb conjugation, syntax adjustment, gender and number agreement and word re-ordering. Once the user has selected the text and set the machine translation process in motion, the program begins to match words of the input text with those stored in its dictionary. Once a match is found, the application brings up a complete record that includes information on possible meanings of the word and its contextual relationship to other words that occur in the same sentence. The time required for the translation depends on the length of the text. A three-page, 750-word document takes about three minutes to render a first draft translation."

At the headquarters of the World Health Organization (WHO) in Geneva, Switzerland, the Computer-assisted Translation and Terminology Unit (CTT) has been a pioneer since 1997 in assessing technical options for using computer-assisted translation (CAT) systems based on translation memory (TM). With such systems, translators can access previous translations from portions of the text; accept, reject or modify them; and add the new translation to the memory, thus enriching it for future reference. By archiving the daily output, the translator helps in building an extensive translation memory and in solving a number of translation issues. Several projects have been under way at the CTT for electronic document archiving and retrieval, bilingual/multilingual text alignment, computer-assisted translation, translation memory and terminology database management, and speech recognition.

The Pan American Health Organization (PAHO) in Washington, D.C. has developed its own machine translation software, as a common work from its own computational linguists, translators, and system programmers. The PAHO Translation Unit has used SPANAM (Spanish to English) from 1980 and ENGSPAN (English to Spanish) from 1985, to process over 25 million words between 1980 and 1998. Staff translators and free-lance translators post-edit the raw output to produce high-quality translations with a 30-50% gain in productivity. The software is available in the LAN (Local Area Network) of PAHO Headquarters, and is regularly used by the staff of technical and administrative units. The software is also available in a number of PAHO field offices, and has been licensed to public and non-profit institutions in the U.S., Latin America, and Spain. The software was later renamed PAHOMTS, and has included new language pairs with Portuguese.

= Comments

# Comments from ZDNN

In "Web Embraces Language Translation", an article published in ZDNN (ZDNetwork News) on 21 July 1998, Martha Stone explained: "Among the new products in the $10 billion language translation business are instant translators for websites, chat rooms, email and corporate intranets. The leading translation firms are mobilizing to seize the opportunities. Such as:

*SYSTRAN has partnered with AltaVista and reports between 500,000 and 600,000 visitors a day on babelfish.altavista.digital.com, and about 1 million translations per day -- ranging from recipes to complete webpages. About 15,000 sites link to babelfish, which can translate to and from French, Italian, German, Spanish and Portuguese. The site plans to add Japanese soon. 'The popularity is simple. With the internet, now there is a way to use U.S. content. All of these contribute to this increasing demand,' said Dimitros Sabatakakis, group CEO of SYSTRAN, speaking from his Paris home.

*Alis technology powers the Los Angeles Times' soon-to-be launched language translation feature on its site. Translations will be available in Spanish and French, and eventually, Japanese. At the click of a mouse, an entire webpage can be translated into the desired language.

*Globalink offers a variety of software and web translation possibilities, including a free email service and software to enable text in chat rooms to be translated.

But while these so-called 'machine' translations are gaining worldwide popularity, company execs admit they're not for every situation. Representatives from Globalink, Alis and SYSTRAN use such phrases as 'not perfect' and 'approximate' when describing the quality of translations, with the caveat that sentences submitted for translation should be simple, grammatically accurate and idiom-free. 'The progress on machine translation is moving at Moore's Law -- every 18 months it's twice as good,' said Vin Crosbie, a web industry analyst in Greenwich, Conn. 'It's not perfect, but some [non-English speaking] people don't realize I'm using translation software.'

With these translations, syntax and word usage suffer, because dictionary-driven databases can't decipher between homonyms -- for example, 'light' (as in the sun or light bulb) and 'light' (the opposite of heavy). Still, human translation would cost between $50 and $60 per webpage, or about 20 cents per word, SYSTRAN's Sabatakakis said. While this may be appropriate for static 'corporate information' pages, the machine translations are free on the web, and often less than $100 for software, depending on the number of translated languages and special features."

# Comments from RALI

Despite the imminent outbreak of a universal translation machine announced at the end of the 1940s, machine translation hasn't produced good translations yet. Pierre Isabelle and Patrick Andries, two scientists from the RALI Laboratory (Laboratory for Applied Research in Computational Linguistics - Laboratoire de Recherche Appliquée en Linguistique Informatique) in Montreal, Quebec, explain the reasons for this failure in "La Traduction Automatique, 50 Ans Après" (Machine Translation, 50 Years Later), an article published in 1998 by Multimédium, a French-language online magazine: "The ultimate goal of building a machine capable of competing with a human translator remains elusive due to slow progress in research. (...) Recent research, based on large collections of texts called corpora -- using either statistical or analogical methods -- has promised to reduce the quantity of manual work required to build a machine translation (MT) system, but can't promise for sure a significant improvement in the quality of machine translation. (...) The use of MT will be more or less restricted to tasks of information assimilation or tasks of text distribution in restricted sub-languages."

According to Yehochua Bar-Hillel's ideas expressed in "The State of Machine Translation", an article published in 1951, Pierre Isabelle and Patrick Andries define three implementation strategies for machine translation: (a) a tool of information assimilation to scan multilingual data and supply rough translation, (b) situations of "restricted language" such as the METEO system which, since 1977, has translated the weather forecasts of the Canadian Ministry of Environment, (c) the human/machine coupling before, during and after the machine translation process, that may not save money if compared to traditional translation.

Pierre Isabelle and Patrick Andries favor "a workstation for the human translator" more than a "robot translator": "Recent research on the probabilist methods showed it was possible to modelize in an efficient way some simple aspects of the translation relationship between two texts. For example, methods were set up to calculate the correct alignment between the text sentences and their translation, that is, to identify the sentence(s) of the source text corresponding to each sentence of the translation. Applied on a large scale, these techniques can use the archives of a translation service to build a translation memory for recycling fragments from previous translations. Such systems are already available on the translation market (IBM Translation Manager II, Trados Translator's Workbench by Trados, RALI TransSearch, etc.) The latest research focuses on models that can automatically set up correspondences at a finer level than the sentence level, i.e. syntagms and words. The results let hope for a bunch of new tools for the human translator, including for the study of terminology, for dictation and translation typing, and for detectors of translation errors."

# Comments from Randy Hobler

In September 1998, Randy Hobler was a consultant in internet marketing at Globalink, after working for IBM, Johnson & Johnson, Burroughs Wellcome, Pepsi, and Heublein. He wrote in an email interview: "We are rapidly reaching the point where highly accurate machine translation of text and speech will be so common as to be embedded in computer platforms, and even in chips in various ways. At that point, and as the growth of the web slows, the accuracy of language translation hits 98% plus, and the saturation of language pairs has covered the vast majority of the market, language transparency (any-language-to-any- language communication) will be too limiting a vision for those selling this technology. The next development will be 'transcultural, transnational transparency', in which other aspects of human communication, commerce and transactions beyond language alone will come into play. For example, gesture has meaning, facial movement has meaning and this varies among societies. The thumb-index finger circle means 'OK' in the United States. In Argentina, it is an obscene gesture.

When the inevitable growth of multimedia, multilingual videoconferencing comes about, it will be necessary to 'visually edit' gestures on the fly. The MIT (Massachusetts Institute of Technology) Media Lab, Microsoft and many others are working on computer recognition of facial expressions, biometric access identification via the face, etc. It won't be any good for a U.S. business person to be making a great point in a web-based multilingual video conference to an Argentinian, having his words translated into perfect Argentinian Spanish if he makes the 'O' gesture at the same time. Computers can intercept this kind of thing and edit them on the fly.

There are thousands of ways in which cultures and countries differ, and most of these are computerizable to change as one goes from one culture to the other. They include laws, customs, business practices, ethics, currency conversions, clothing size differences, metric versus English system differences, etc. Enterprising companies will be capturing and programming these differences and selling products and services to help the peoples of the world communicate better. Once this kind of thing is widespread, it will truly contribute to international understanding."

= Machine translation R&D

Here is an overview of the work of four research centers, in Quebec (RALI Laboratory), California (Natural Language Group), Switzerland (ISSCO) and Japan (UNDL Foundation).

# RALI Laboratory

In Montreal, Quebec, the RALI Laboratory (Laboratory of Applied Research in Computational Linguistics - Laboratoire de Recherche Appliquée en Linguistique Informatique) has worked in automatic text alignment, automatic text generation, automatic reaccentuation, language identification, and finite state transducers. RALI produces the "TransX family" of what it calls "a new generation" of translation support tools (TransType, TransTalk, TransCheck, and TransSearch), which are based on probabilistic translation models that automatically calculate correspondences between the text produced by a translator and the original text from the source language.

As explained on RALI's website in 1998: "(a) TransType speeds up the keying-in of a translation by anticipating a translator's choices and criticizing them when appropriate. In proposing its suggestions, TransType takes into account both the source text and the partial translation that the translator has already produced. (b) TransTalk is an automatic dictation system that makes use of a probabilistic translation model in order to improve the performance of its voice recognition model. (c) TransCheck automatically detects certain types of translation errors by verifying that the correspondences between the segments of a draft and the segments of the source text respect well- known properties of a good translation. (d) TransSearch allows translators to search databases of pre-existing translations in order to find ready-made solutions to all sorts of translation problems. In order to produce the required databases, the translations and the source language texts must first be aligned."

# Natural Language Group

The Natural Language Group (NLG) at the Information Sciences Institute (ISI) of the University of Southern California (USC) has been involved in various aspects of computational/natural language processing: machine translation, automated text summarization, multilingual verb access and text management, development of large concept taxonomies (ontologies), discourse and text generation, construction of large lexicons for various languages, and multimedia communication.

Eduard Hovy, head of the Natural Language Group, explained in August 1998: "People will write their own language for several reasons -- convenience, secrecy, and local applicability -- but that does not mean that other people are not interested in reading what they have to say! This is especially true for companies involved in technology watch (say, a computer company that wants to know, daily, all the Japanese newspaper and other articles that pertain to what they make) or some Government Intelligence agencies (the people who provide the most up- to-date information for use by your government officials in making policy, etc.). One of the main problems faced by these kinds of people is the flood of information, so they tend to hire 'weak' bilinguals who can rapidly scan incoming text and throw out what is not relevant, giving the relevant stuff to professional translators. Obviously, a combination of SUM (automated text summarization) and MT (machine translation) will help here; since MT is slow, it helps if you can do SUM in the foreign language, and then just do a quick and dirty MT on the result, allowing either a human or an automated IR-based text classifier to decide whether to keep or reject the article. For these kinds of reasons, the U.S. Government has over the past five years been funding research in MT, SUM, and IR (information retrieval), and is interested in starting a new program of research in Multilingual IR. This way you will be able to one day open Netscape or Explorer or the like, type in your query in (say) English, and have the engine return texts in *all* the languages of the world. You will have them clustered by subarea, summarized by cluster, and the foreign summaries translated, all the kinds of things that you would like to have."

Eduard Hovy added in August 1999: "Over the past 12 months I have been contacted by a surprising number of new information technology (IT) companies and startups. Most of them plan to offer some variant of electronic commerce (online shopping, bartering, information gathering, etc.). Given the rather poor performance of current non-research level natural language processing technology (when is the last time you actually easily and accurately found a correct answer to a question to the web, without having to spend too much time sifting through irrelevant information?), this is a bit surprising. But I think everyone feels that the new developments in automated text summarization, question analysis, and so on, are going to make a significant difference. I hope so!--but the level of performance is not available yet.

It seems to me that we will not get a big breakthrough, but we will get a somewhat acceptable level of performance, and then see slow but sure incremental improvement. The reason is that it is very hard to make your computer really 'understand' what you mean -- this requires us to build into the computer a network of 'concepts' and their interrelationships that (at some level) mirror those in your own mind, at least in the subjects areas of interest. The surface (word) level is not adequate -- when you type in 'capital of Switzerland', current systems have no way of knowing whether you mean 'capital city' or 'financial capital'. Yet the vast majority of people would choose the former reading, based on phrasing and on knowledge about what kinds of things one is likely to ask the web, and in what way. Several projects are now building, or proposing to build, such large 'concept' networks. This is not something one can do in two years, and not something that has a correct result. We have to develop both the network and the techniques for building it semi-automatically and self-adaptively. This is a big challenge."

Eduard Hovy added in September 2000: "I see a continued increase in small companies using language technology in one way or another: either to provide search, or translation, or reports, or some other communication function. The number of niches in which language technology can be applied continues to surprise me: from stock reports and updates to business-to-business communications to marketing...

With regard to research, the main breakthrough I see was led by a colleague at ISI (I am proud to say), Kevin Knight. A team of scientists and students last summer at Johns Hopkins University in Maryland developed a faster and otherwise improved version of a method originally developed (and kept proprietary) by IBM about 12 years ago. This method allows one to create a machine translation (MT) system automatically, as long as one gives it enough bilingual text. Essentially the method finds all correspondences in words and word positions across the two languages and then builds up large tables of rules for what gets translated to what, and how it is phrased.

Although the output quality is still low -- no-one would consider this a final product, and no-one would use the translated output as is -- the team built a (low-quality) Chinese-to-English MT system in 24 hours. That is a phenomenal feat -- this has never been done before. (Of course, say the critics: you need something like 3 million sentence pairs, which you can only get from the parliaments of Canada, Hong Kong, or other bilingual countries; and of course, they say, the quality is low. But the fact is that more bilingual and semi-equivalent text is becoming available online every day, and the quality will keep improving to at least the current levels of MT engines built by hand. Of that I am certain.)