When the globalization organization at Sun Microsystems, Inc., investigated automation of the translation process, the organization concluded that Sun could maximize the text that was machine translatable. Translation memory tools and translation processes were continuously improving the localization process of Sun and of its vendors. However, a larger increase in the translation production cycle could only be achieved through further automation, particularly machine translation. The globalization organization concluded that machine translation would support the translation of a constantly growing amount of material efficiently and effectively. Sun could then provide a significant increase in the throughput volume and speed for delivering localized content.

The idea of writing for machine translation was introduced at Sun Microsystems in 1999. The argument that clear and unambiguous text would be easier to understand and to translate for humans as well as machines was easily accepted. However, the effect on the writing process, and the means by which these changes would be enforced, were very challenging. The addition of new rules to writing guidelines was the first concern. The second concern was the addition to the publication process of the use of an application that could help writers identify translation problems and rewrite sentences accordingly. These adjustments to the writing process had to be carefully introduced.

Three years later, Sun has a success story. The Solaris™ Technical Publications group is now the first publishing group at Sun Microsystems that improves translatability in their daily authoring with customized machine-translation software. The technical writers for Solaris software check the translatability of Solaris documentation with the help of an application that was developed and customized by the Institute of Applied Information Sciences (IAI). IAI based its customization on translatability guidelines that were developed by Sun.

Extensive comparative analysis and research of companies that offer translatability checkers or controlled language systems led to a pilot project with the Controlled Language Authoring Technology (CLAT). The CLAT system, created by IAI in Saarbrücken, Germany, and customized for Sun Microsystems, is now called Sun™ Proof.

A focus group worked together to customize and introduce the Sun Proof application for use at Sun. The group included writers, editors, and managers from the Solaris Technical Publications group along with members of the translation technologies team, Sun Microsystems Laboratories, and the IAI team. Close cooperation, well-defined roles, and shared responsibility between some key organizations inside and outside of Sun Microsystems led to the successful deployment of the Sun Proof translatability guidelines. As a result, Sun is now on its way to successfully introducing machine translation into the Sun Microsystems product development process.

Development of a Translatability-Checking Application

Key to the success of the translatability-checking application was the focus group's set of requirements for implementing a translatability-checking application at Sun Microsystems.

The translatability-checking application had to satisfy the following primary requirements:

• Support Sun Proof translatability guidelines (Sun should be able to customize rules)
• Run on a UNIX platform
• Support SGML and the editing tool used by Solaris technical writers
• Interpret SGML tags
• Support interactive and batch mode checking
• Perform terminology checking
• Allow the possibility to overrule a translatability guideline

The focus group of writers, which included eight technical publication writers, conducted numerous tests on the various releases of the Sun Proof application over a two-year period. The primary purpose of their testing was to evaluate the effectiveness of the candidate set of translatability guidelines for improving the translatability of the Solaris documentation. The writers also provided feedback on the usability of Sun Proof and how translatability checking affected the English source document. The writers chose a document of no more than 10 pages, ran Sun Proof on the text, and edited the document by following the translatability guidelines. The writers were told to track their corrections in a test log, and if they decided to ignore a translatability critique, to give an explanation. The writers then submitted the original chapter and the rewritten chapter along with their test logs to the editors.

The editors tested Sun Proof to evaluate writer accuracy in following the translatability guidelines, determine how Sun Proof affected the English source document, and establish whether revised texts would reduce editing time. The editors edited both the "before" and "after" text while tracking usual errors that they would highlight in a review and any Sun Proof guideline violations. The editors then reviewed the writer test log for explanations on changes to sentences and reasons for ignoring Sun Proof critiques.

Analyzing the writer and editor test logs showed the Localization group which rules were violated most often, which rules the application handled well, and which rules needed improvement. As a result, many of the original guidelines were replaced, modified, or rewritten to give clearer instructions. Wherever possible, requests for interface and usability enhancements were fulfilled and the application was customized to fit the existing writing and publication process.

The feedback from the testers was very positive. Overall, the testers felt the guidelines were easy to follow, were logical, and even improved their writing style. The editors were also very supportive. These individuals found that many of the corrections they would usually suggest were caught in the Sun Proof review process. Editing of revised texts did not seem to be much faster. However, the editors needed to focus less on minor, repetitive errors that were now caught by Sun Proof. Over time, the editors became so familiar with the Sun Proof rules, that they would incorporate translatability edits in texts that were not checked by Sun Proof.

For the translation portion of the pilot project, the Sun Localization group worked with students of translation and interpretation at the Monterey Institute of International Studies. Translators of German, Spanish, Japanese, and Chinese participated in the study. The translators were given two sets of files for analysis: the original documentation and the rewrites that were based on Sun Proof feedback to writers. The translators used third-party machine translation applications for the translation of the original and revised texts. The goal was to ensure the translation would be unbiased and unaffected by repeat translations of similar material. The translators analyzed the translation quality of the machine translated texts based on a translation quality metric. This metric consisted of the following six categories of translation errors:

• Mistranslation of a term
• Syntactic error
• Word structure or agreement error
• Non-translation of a term
• An untranslated term
• Punctuation error

A comparison was then made between the translation quality of the original texts and the quality of the rewritten texts.

The results ranged from minor improvements at the beginning of the project in January 2000 to major improvements by January 2002. The increase in translation quality was a result not only of refining the translatability guidelines, but also of leveraging the writers' experience in rewriting sentences.

The translation improvements were very convincing. As a result, the Sun translation technology team, along with the technical publications group, decided to proceed with introducing Sun Proof into the documentation production process.

Translatability Checker: Sun Proof

The production version of Sun Proof was introduced to the Solaris writers in April 2002. The final set of translatability guidelines was decided by the focus team of writers and editors and the translators at the Monterey Institute. The Sun Editorial Forum then approved the guidelines. The Editorial Forum is a committee of writers and editors who establish Sun style guidelines and publish the Sun Editorial Style Guide. Approval from the Editorial Forum was granted for the content and wording of the guidelines, and for the explanations and examples incorporated in the Sun Proof application.

The Sun Proof translatability guidelines can be divided into three categories: style guidelines, grammar rules, and terminology. Approximately 30 style guidelines are in Sun Proof 1.0. All of the guidelines promote simplicity, clarity, or both. One of the most effective guidelines is the rule that limits sentence length to 25 words. Sentences that exceed that length tend to have too many dependent clauses and present a translation challenge to human translators as well as machine translation. Here is an example of a long sentence that was rewritten by using Sun Proof guidelines. The rewrite dramatically improved the machine translation.

Only a few grammar rules influenced translatability throughout all tests. The final set was chosen on the basis of two qualifiers. First, a rule was included if a violation could result in a meaning shift in the original and the translation. Second, a rule was included if a violation could lead to a misparse in the machine translation application. One example of a grammar error that is flagged incorrect is subject and verb agreement.

Correct and consistent usage of terminology is known to have a great effect on translation. As a first step, the Sun Proof project team decided to focus on terms that should not be used or are less preferred variants of terms. In most cases, the writers were already familiar with these terms from the Sun Editorial Style Guide. However, automating the detection of "illegal" terms and providing suggested rewrites was believed to promote more consistency. Automation could eliminate the need to look up the offending term in the style guide. An example of a word that would be flagged is the verb "may." Writers are reminded to use this word only when granting permission and not as a replacement for "can" or "might." Machine translation of the word "may" in the latter meaning always produces a shift in meaning.

Implementation

Some key factors contributed greatly to the successful implementation of Sun Proof in production. The first factor was the direct involvement of the writer and editor community in customizing the application. Other factors were the integration into the existing work environment and targeted training to support the integration.

The Solaris application team integrated Sun Proof with the existing SGML editor to provide a familiar, user-friendly access to the translatability application. Writers are now able to check a document that they are working on in their SGML editor by clicking on a link to Sun Proof. After the text has been reviewed for translatability problems, the revised document is saved directly to the SGML editor. The writer can then continue with the writing process.

For those writers who prefer to complete their work before checking it, the application team and IAI created a batch-mode checking option. After a short processing time, this option provides the writer with a PDF file that highlights all the Sun Proof errors in the text. Additionally, the writer receives a proof sheet that lists the errors and their location in the file.

Finally, for writers who do not work in the standard SGML editor, the option exists to run Sun Proof as a standalone application. This option is particularly popular with writers who work remotely.

Sun Proof training would also become an integral part of the application's successful release into production. The Sun Localization team gave numerous presentations and demonstrations before implementing Sun Proof. As Sun Proof was implemented, everyone in the writing community knew about the project and the concept when they needed to learn the application's features and start using them daily. Sun Proof classes were held across the country at various Sun facilities. Groups of six to eight people intensively trained in three hour classes on how to use Sun Proof. These students focused on rewriting sentences according to the translatability guidelines, and writing with translation requirements.

The experience with the pilot project predicted a learning curve in using Sun Proof effectively and translatability results that might not be completely satisfactory at first. Translators indicated that some sentences that had been rewritten as advised by Sun Proof did not improve in clarity. Further, in some rare cases, some sentences became even more ambiguous after revisions. These sentences were submitted to the Localization group for analysis. The Localization group concluded that the guidelines that the writers seemed to be having the most difficulties with were long, complex sentences. These were the same guidelines that had demonstrated the most improvement to translation when sentences were rewritten correctly. The writers had as much difficulty as translators did in interpreting and clarifying complex sentences. This finding indicated again to make these sentences unambiguous from the start. Consequently, writing workshops were organized that focused almost exclusively on how to rewrite long or complex sentences. These workshops were offered to all new Sun Proof users at Sun.

Sun Proof: Status and Future Direction

Sun Proof is now used by over 40 Solaris writers. So far, Solaris writers have used Sun Proof on over 30 Solaris books, which amounts to approximately 9,000 pages of documentation. This documentation will be significantly easier to translate because of the number of sentences that were rewritten to satisfy Sun Proof guidelines. More importantly, the time to translate the next release of the Solaris documentation will be reduced as a result of machine translation. Because the time to translate the Solaris documentation will be reduced, Solaris writers will have their content-final date extended. This date is more aligned with the stabilization of the Solaris product, ensuring the accuracy of the documentation.

More and more, Sun product writers are being trained to use Sun Proof on their writing efforts. Further analysis will be conducted to improve translatability guidelines when a larger corpus of "before" and "after" documentation will be available from the current Solaris development release. Finally, with such a large corpus available, various analyses will be conducted to quantify actual time and cost savings.

Conclusion

As a result of introducing Sun Proof into the documentation development process at Sun Microsystems, the following benefits have been gained:

• Sun is now one year away from introducing machine translation into the Sun product development process.
• Human translation time is reduced because the troublesome sentences have been rewritten to make the sentences easier to translate.
• Because sentences are easier to translate, the translation accuracy has improved.
• Writers can delay their documentation delivery to the localization groups, which gives the writer more time to accurately document the product when the product has stabilized.
• Rewriting troublesome sentences to Sun Proof guidelines has made all the Solaris documentation more consistent in sentence structure.
• Rewriting troublesome sentences has reduced ambiguity in the documentation, which has improved the readability of the English documentation.

Jennifer Wells Akis is Project Manager in Customer Network Platform, eProductization: Translation Technologies at Sun Microsystems.

William R. Sisson is Senior Technical Writer in Solaris Information Products: Systems and Network Administration at Sun Microsystems.

The authors would like to thank Susanne Andersson, Bob Kuhns, and PJ Schemenaur for their review of this article and for many helpful and insightful suggestions.