Eduard Hovy
Perhaps within a decade, MT (machine translation) will appear in your life in several ways--translating telephones, multilingual E-mail, and machines that scan and translate letters and articles written in foreign languages. You may be buying toys over the phone from a sales agent in Japan with the telephone doing the translating. And when you travel to a foreign country, you'll be able to get the same bargain rates that the natives do with your trusty PET (Portable Electronic Translator).
Several technologies will make such scenarios possible: automated speech recognition, speech generation, OCR (optical character recognition), and machine translation. Although PETs are years away, intermediate tools are already available, either as research prototypes or as commercial products.
From "Mushi-Mushi" to "Hello"
One of the most complex component technologies is speech recognition. After about 30 years of research, systems can recognize limited amounts of speech without first having been trained on specific voice patterns. Systems have error rates of only a few percent when the domain (i.e., a vocabulary specific to a certain environment or discipline) is limited to a few thousand words.
Most speech-translation efforts are taking place in Japan. Researchers at the ATR Laboratories (Advanced Telecommunications Research Institute International, Kyoto, Japan) have been working on a translating- telephone project for about six years. Working in English and Japanese, the prototype system, SL-TRANS, will translate inquiries about conference registrations. The company's goal is for the system, using a vocabulary of 1500 words, to be able to translate anyone's speech without the system being tailored to specific voice patterns.
Speech-recognition systems combine hardware (e.g., microphones) with statistically based software to match incoming phonemes (i.e., sounds) against stored phoneme patterns and to produce strings of text. The best experimental systems using standard workstations operate at well over 90 percent accuracy on domains of a few thousand words. An NEC system can recognize a 5000-word vocabulary of isolated words or a 1500- word vocabulary of fluid speech.
A consortium of 10 universities and companies in Germany is planning an ambitious speech-translation project. Personnel associated with the Verbmobil project will build a portable device to help translate languages for business discussions. Because it's easier to understand a foreign language than to speak one, the project personnel say, the system will use English as a kind of interlingua.
The Bibliothque Nationale
Multilingual information-retrieval systems are closer to becoming commercially available than speech-recognition systems are. With retrieval systems, you'll be able to get information on your favorite topic by giving a collection of keywords or a few relevant paragraphs.
Researchers at Fujitsu Laboratories (Tokyo, Japan) have built a multilingual information-retrieval system that translates query words in German or English into Japanese. It retrieves relevant articles; translates their titles into German or English; and, on request, translates the article as well.
E-Mail from A to Z
Computing environments are becoming increasingly dependent on networks. Communications are more electronic and less paper-based than they were in the past. Thus, people have a greater need for inexpensive technology that offers browsing-quality translation assistance of E-mail messages.
Although most of the technology for this application already exists, no products are on the market yet. Fujitsu has demonstrated a prototype, and Dragon Systems (Boston, MA) is developing a PC-based product that will handle Japanese, Spanish, and English translations.
Future Applications
In a competition sponsored by Apple's Advanced Technologies Group, students at Carnegie Mellon University's School of Design offered several innovative ideas. One prizewinner was the Illumina, an MT machine that stands in an office or library next to the printer and photocopy machine. Combining optical-scanning and OCR technology with personal computer-based MT, the Illumina translates whatever text you place on its faceplate. Given the high character-recognition rate available in commercial systems, this type of system may start appearing as soon as more general-domain MT systems develop out of the prototype stage--perhaps by the end of the decade.
Carnegie Mellon students have also developed the Passport, a hand-held word-translation device that uses photography to input text (see photo A). For example, on your travels, you spot graffiti on a wall. After photographing the slogan onto the Passport's display, you circle the words you're interested in, and the Passport's bilingual dictionary chip provides a translation on an LCD. Granted, for a few hundred dollars you can buy systems that will translate the foreign words that you type in, with a separate chip for each language pair, but try typing in a few Chinese characters.
A third Carnegie Mellon project addresses the needs of hearing- impaired people. The Signspeak System is intended to translate American Sign Language and spoken English (see photo B). Its design calls for a glove with sensors and an attachment to track arm position and hand motion. These sensors feed their signals to a personal computer, which converts them into English words and uses voice-synthesis software to convert the reply to text, which is displayed on the computer's screen.
Several neural-network and connectionist project developers are also working toward future MT applications. Carnegie Mellon and the University of Karlsrule in Germany are jointly developing the C-Star System. In this system, trained connectionist networks recognize English text, create internal data structure patterns, and generate Chinese text as output. A set of neural-network programs called ANN (Architecture for Neural Networks), developed at the Huntsville Achievement School (Huntsville, AL), follows a similar approach with English and Russian.
These applications are but the vanguard of the MT revolution. More have yet to come. Over the past few years, fax machines have quietly changed our lives. In a few decades, automated MT systems will be doing the same. Thanks to the innovative use of MT and related technology, speech and writing in a multitude of languages will become easier to find, read, and understand.
Photograph A: The Passport combines digital photography and character-recognition technologies with bilingual dictionaries on chips. Using this hand-held device, you can capture an image of something you'd like translated and see the translation of the words or symbols. (Courtesy of Carnegie Mellon University)
Photograph B: To help hearing-impaired people communicate with those who don't use American Sign Language, students at Carnegie Mellon developed this prototype translator. The design uses gloves with sensors to send hand- and arm-movement information to a personal computer. Software converts the signals into English words, which are synthesized and output over a loudspeaker. (Courtesy of Carnegie Mellon University)
Eduard Hovy is a project leader at the Information Sciences Institute of the University of Southern California in Los Angeles. He is the author and co-editor of two books on automated language generation. You can reach him on BIX c/o "editors" or on the Internet at hovy@isi.edu.
Per comentaris i observacions sobre el servidor, poseu-vos en contacte amb l'Administrador WWW