Amikai: Best-Of-Breed MT Engine
Amikai was founded by three Stanford University graduates in 1999 in San Francisco and opened its Tokyo office in 2000. The word "Amikai" comes from the Latin 'amicus' (friend) and the Japanese word 'sekai' for 'world'. The company develops Machine Translation (MT) applications to facilitate Web-based multilingual communication. The Japan Industrial Journal in April 16, 2001 described the company as possessing the three fundamentals for a successful Internet business: human capital, technical know-how and financial backing.
What specifically brought Amikai to the attention of LISA's Asia-Pacific Editor, Minako O'Hagan, was the fact that a number of separate Japanese online translation sites return exactly the same results. This eventually led to the discovery that these sites were indeed all powered by Amikai. As of February 2003, Amikai has been implemented as the default translation engine by Excite Japan, @nifty.global gate, Cafeglobe, Infoseek (Japan) and Lycos (Japan). This is a significant achievement, given that the market had previously been dominated by MT systems developed by Japanese companies.
LISA caught up with Amikai's Director of Computational Linguistics, Dr Raymond Flournoy, to learn about the company's approach to MT-based language solutions for communicating on the Internet, from both the business and technical perspectives.
Minako O'Hagan: To kick off this interview, could you give us some figures on how many hits Amikai translation engines are typically receiving in a day (if possible, according to different language pairs)? And, do you have a mechanism to incorporate user feedback on translation quality to improve your system?
Raymond Flournoy: Amikai serves well over a million translation requests per day across 29 uni-directional language pairs (14 bi-directional pairs plus one uni-directional pair), with the majority of traffic consisting of translation between Japanese and English. Currently, the Amikai translation system does not include any automatic methods for incorporating user feedback. However, we actively solicit user comments and conduct client surveys in order to improve our translation products through regular product updates.Do any of the translation sites that use Amikai MT systems charge for the translation? Do you have usage statistics for such services (in comparison with free usage)? How are these services received by users?
In addition to the Amikai-powered translation services that are provided for free on a number of portal and ISP Web pages, Amikai provides a for-pay translation service, primarily for business clients, called Amikai Enterprise. We do not currently have usage statistics that we can release. However, the pay service reaches a significantly different audience than the free service. The portal and ISP translation services include translated browsers, translated chat and simple text translations. Amikai Enterprise is integrated with common desktop applications such as Microsoft Office and Outlook, and the translations are optimized for language as used in a business setting.Amikai is a language solutions company that produces MT applications designed for the Internet (Web) and your revenue comes from the support and maintenance provided with Amikai systems. But, you are different from MT developers proper since you utilize third-party MT systems. Do you envisage Amikai continuing on this basis or will it, at some stage, implement its own research and development function to create an entirely new MT system?
Amikai's products incorporate a mixture of in-house, as well as licensed, technologies. While the core translation engines are currently licensed from other companies, Amikai continues to develop its own technologies. Our current emphasis, and the area where we feel we can make the most significant contribution, is to increase the usability and quality of currently available translation products. We do this by improving translation quality for specific translation applications, by innovating interfaces and links to other software, and by improving the stability of translation software.Amikai has been able to attract some of the major Japanese Web sites to power their translation utilities and, therefore, has gained a reputation as an MT system which adds practical value. Could you share with us some of your customer feedback on the reason(s) why they chose Amikai? Also, from the language engineering point of view, what are the superior features of Amikai systems?
Amikai has become a leader in the field of on-line machine translation because we provide high-quality translations that are geared to specific communication usages. For example, our translated chat rooms are optimized for chat-specific language, while our translated Web browsers consult special dictionaries to improve coverage of the language common to Web pages.
In addition, we have built a server farm that guarantees stable access to fast translations, regardless of which of the 29 language pair-directions one is translating. Finally, Amikai also places a special emphasis on the user experience, and we feel this pays off in a more useful and enjoyable product. This conclusion has been supported by the feedback from clients and end-users who repeatedly tell us that we have a higher-quality translation product than most other services they have tested.Amikai offers a range of products that includes AmiWeb, AmiText and AmiChat to cater to different needs that come from different modes of communication. How do you design MT systems to be optimized for a given mode? Also, are there any differences in the design for different language pairs, e.g., AmiWeb for English/Japanese vs. AmiWeb for English/Spanish?
When Amikai designs an application, a number of factors are taken into consideration. First, we consider how the tool is used – whether the translated inputs are short or long, the time-sensitivity of the translation requests and the appropriateness of MT quality. Next, we study the basic composition of the language to be translated – the prototypical vocabulary, commonly repeated phrases and recurrent sentence structures. At this stage, the analysis is language-specific, and separate processing and lexicography must be performed for each source language.Amikai is described as a "best of breed" MT provider. What does this mean?
It means that for each language pair that we support we provide access to the (single) best translation engine available on the market. We run tests on available engines to determine the best overall for our purposes, and then license that technology.As I understand it, Amikai has worked on an MT system architecture based on a dynamic selection of the best output from existing commercially available MT systems, using a statistical language modelling technique. This sounds like an intriguing design approach. Can you explain how it works? Is this technique used currently for Amikai systems?
The idea of dynamically choosing among a variety of translation engines is just one innovation that Amikai has experimented with in our quest to improve translation quality. The program simultaneously sends the same translation request to a number of engines, and then matches each result against a statistical language model to see which result best fits language patterns the program has "seen" in the past. While this feature is not part of our current commercial offerings, it is an example of the innovation that we bring to the field of MT.Recently, statistical and corpus-based techniques seem to have become fashionable in MT design and with tools such as translation memory systems. Is the dynamic selection approach inspired by these trends? Where did this design idea come from?
This approach was inspired by the work of Kevin Knight, a professor at the University of Southern California. Professor Knight has used language models as a component of his statistical MT algorithms, and Amikai has adapted this part of his approach to the task of evaluating MT engine outputs. Statistical language modelling is a very flexible and promising tool, and we are excited about this possible application.In many ways, MT was revitalized by the emergence of the Internet to satisfy the latter's need to allow communication in any language in real-time cheaply, but not necessarily with precise accuracy. How do you see the role of human translators, as opposed to machine translation, in this environment?
Amikai believes strongly that MT is not a replacement for Human Translation (HT), but rather, MT is appropriate and useful in situations where constraints such as time and cost make HT unfeasible. Those situations include communication where the turnaround time must be nearly instantaneous (such as chat), where the amount of text is prohibitively large (such as large datastores), and where the text is continuously updated (such as the Web).The localization industry strives to perform high-quality localization within specific budgets and timeframes imposed by clients. To this end, the industry is leveraging technology, but combining it with human talent. By comparison, WebMT provides a just-in-time automatic localization solution for those Web sites that are not provided in the end-user's locale. How do you see this juxtaposition of carefully localized Web sites by human sweat (!) assisted by computer, versus those that are not localized, and rely on an ad hoc MT solution by the user of a given Web site?
Amikai feels that MT and HT are highly complementary. HT provides the highest quality of translation for sites which are very brand-sensitive or which require high accuracy (such as legal information). MT allows users to access sites that have not been translated by humans and, which otherwise, would remain inaccessible due to the language barrier.Amikai is mostly known for facilitating B2C (business-to-commerce) interactions via Web sites, but it is also reported to be moving into the B2B (business-to-business) space. What is the B2B scenario with MT, i.e., Amikai products?
Although most of our translations have been performed through Web portals or ISP's, we do not consider those to be B2C sales. The application user is the portal visitor or ISP customer, but Amikai's customer is the portal or ISP itself. We are now also expanding into enterprise-oriented translations. To Amikai, both of these scenarios are B2B. However, in the first case, the primary user of the MT service is outside of the client company, while in the enterprise case, the primary user is within the client company.In relation to the B2B application of MT, what would be the scenario with enterprise-wide MT applications by multinationals that want to facilitate interaction among their international staff across language barriers?
There are many uses for MT in international business. Translated chat is useful for collaboration among linguistically diverse teams. Translated Web browsers facilitate corporate research across language barriers. Translated email eases communication both within and outside of an international business. Translated Word documents, PowerPoint presentations, Excel spreadsheets and Acrobat documents allow access to corporate information, regardless of the source language.From the start, Amikai seems to have made it clear that Japan is a very important market for the company. Traditionally, translation requirements in the Japanese market have generally concentrated on Japanese-English. But recently, an emerging need has surfaced for Japanese to be translated into other Asian languages such as Chinese and Korean and vice versa. Is this something Amikai is working on, especially since it already has Korean within its range of languages?
Yes, Amikai is actively working on expanding and improving its offerings in Asian languages other than Japanese, e.g., Korean and Simplified Chinese.What is Amikai's next big thing?
At Amikai, we are very excited about the launch of the Amikai Enterprise translation solution, and we are eagerly watching its adoption in international corporations.The Internet has made MT widely visible to the general public. Within a very short timeframe, translation has come to be accepted as one of the standard functionalities of the Internet. Authors of the 1966 ALPAC report would be astonished to see this development with MT. This illustrates how difficult it is to predict the future. Having said that, what is your view on the future of MT?
MT is a technology just now coming into its own commercially. Like any tool, MT has appropriate uses and inappropriate uses, and at Amikai, we believe that as users develop a better sense for the usefulness of MT, more and better commercial applications will be developed. There is still room for great advances in quality, but in addition, engines will need to expand beyond the current sentential unit to consider longer contexts when performing translations. In addition, we believe that there is much untapped potential in the types of MT applications that are available, including ones that better integrate MT and HT. All in all, Amikai is excited to be at the cutting edge of MT applications, and we see a bright future for this technology.
Raymond Flournoy, Ph.D. received his B.A. from Harvard College, and after studying at Kyoto University under a Monbusho Fellowship, proceeded to Stanford University's Department of Computer Science. He received his Ph.D. in 2000, and since then has been the Director of Computational Linguistics at Amikai, Inc. in San Francisco, CA, USA. Fluent in Japanese, his primary areas of interest are Machine Translation and Cross-Language Information Retrieval.
Reprinted by permission from the Globalization Insider,
18 March 2003, Volume XII, Issue 1.5.
Copyright the Localization Industry Standards Association
(Globalization Insider: www.localization.org, LISA: www.lisa.org)
and S.M.P. Marketing Sarl (SMP) 2004
Please see some ads as well as other content from TranslationDirectory.com: