24/7 English to Japanese Alert Translation for a Global Security & Intelligence Company
A global security and intelligence company (hereinafter referred to as “GSIC”) conducted a competitive bid for a language services provider who could provide live, 24/7 English to Japanese translations without interruption and without delay. The GSIC’s mission was to deliver time-sensitive and actionable information, on threat intelligence and response services, to enable its corporate and government clients to mitigate global risks. The contract involved the translation of travel briefs, health briefs, and other content involving 2 million+ words. However, the most challenging task was to build a continuous translation production capability to support the daily and spontaneous, critical alerts distributed by GSIC, as daily global events dictated.
From among eleven vendors, Compass Languages (“Compass”) was awarded the contract in late June. Compass had only three months to prepare for the launch on 1 October of that year. This was a perpetual contract and this case study explains the challenges, solutions, and accomplishments over time, during the contract.
Initial Technical Challenges
Linguists: A team of sixteen linguists was initially selected to begin the translation of filed reports into Japanese – a 1.2 million word exercise to build up the Translation Memory (hereinafter referred to as “TM”) necessary to keep alert translation costs affordable. However, the real challenge was to recruit Japanese native speakers who lived in targeted time zones that could support the 8-hour shifts required for continuous alert translations by the October 1 launch date.
Workflow Platform: The GSIC’s high quality demands on Japanese translations meant that Compass would employ its full ISO 9001 process of two independent human translators on every alert translation. The 24/7 activity required the establishment of an automated workflow process where PMBots would replace a human project manager. (PMBots are software systems that replace the actions/decisions made by a human project manager so that 24/7 operation can be managed without the time, resources, and limitations of human project managers.) The PMBots would detect when new content was available (every 10 seconds), process the English content through the TM application, handoff the data to the linguists, execute quality assurance tests, and then deliver the Japanese translated content to the client’s CMS.
Turnaround: The GSIC originally wanted translations within 30-45 minutes, but with the standards for high-quality translations and the average word count of 300 words, the company accepted 90 minutes for delivery of these average length alerts. The turnaround requirement created a need for additional PMBot sophistication in order to assure the true “Standby” status of two individuals every minute of every day.
Zero Interruption Tolerance: The GSIC’s marketplace position required that Compass create redundant systems such that translation service would never be interrupted by any event for more than 60 minutes. This disaster recovery system involved many layers and covered all conceivable outages.
Compass chose Kilgray’s MemoQ Software and Workflow Platform for management of all alert projects and the translation memory creation coming from the cumulative translation effort. A former Kilgray engineer was hired to build the PMBot capability. The automation of file handling, TM application, vendor selection, quality assurance, and file delivery was completed within 6 weeks.
Additional aspects of the decision-making process for the PMBot took longer to refine, such as:
- What to do when a standby linguist did not respond
- How to handle update alerts that entered system while first version was still in translation
- How to capture statistics of each alert (7,000 alerts per year) for invoicing and PO generation
- How to warn linguists that too much time was spent on any given task
- How to monitor file progression and know that an alert translation was “stuck”
- How to manage the various QA errors and know the difference between minor and major errors
For three weeks during the October launch, Compass employees worked around the clock monitoring all activities of the translation process and manually over-riding the PMBots when a situation could not be resolved. By late October, the PMBots were fully operational and managing 99.9% of all decisions.
In order to meet the requirement for 24/7/365 continuous translation, Compass trained 20-25 EN>JP linguists who lived in targeted locations around the world. Two linguists were paid to be on standby for each 8-hour shift. Compass established a North American zone, a European zone, and Japanese zone. At least 5 linguists in each zone were needed to be available each month. All others were available if scheduled linguists were not available or if the workload became overwhelming at any period of time. During certain world events, the GSIC could be writing 5-6 alerts per hour. Most alerts were in the range of 200-350 words each. Approximately 10% of the alerts were over 600 words.
To accommodate the requirement of zero interruption > 60 minutes, Compass created a redundant PMBot system with (3) Bots working on independent machines and coordinating their activities. Each Bot was completely capable of managing the entire translation process. Backup systems were incorporated for electrical or internet service interruptions. Warning systems utilizing text and Skype messaging were established for key Compass staff to respond when interruptions occurred.
After 2 years of performance, Compass began developing a Machine Translation engine to help GSIC reduce their total translation costs. Not impressed with the results, we transitioned to a Neural Machine Translation (hereinafter referred to as “NMT”) solution. The more traditional machine translation engine builds were not as interesting because the domain was too broad to accomplish the accuracy the company needed. With over 4 million words translated, and the learning capabilities inherent within NMT engines, Compass decided to train an engine to reach our standards for first-pass accuracy. With a design-to-cost goal in mind, Compass established the post-editing function to be in the 25-30 minute range for the typical alert size of 300 words. This would achieve the 40% reduction in translation cost we were trying to achieve. In addition, the NMT methodology with human post-editing and software QA tools would allow us to lower the average turnaround time from 90 minutes to under 45 minutes.
For over three months, Compass performed this NMT/PE process on Warning alerts only (no Critical alerts) and used a second human linguist to review the work of the post-editor before releasing the translation for live viewing. Compass monitored the time required by the second reviewer and the “edit-distance” of perfecting the translation. After approximately 90 days, the GSIC was satisfied with the results accomplished, and were willing to drop the review function. The cost reduction of 40% was deemed worth the slight increase in error rate.
Executive Summary of Accomplishments
- Compass was operational, within 90 days of contract award, providing real-time English to Japanese (hereinafter referred to as “EN>JP”) alert translations.
- Unbeknownst to Compass, the GSIC was demonstrating this technology “live” at a Trade Show in Europe on October 1.
- Compass completed the translation of all 1.2 million words contained in Intelligence briefs within 4 months.
- In 40 months of continuous translation, Compass experienced only two service interruptions. Both interruptions were overcome within 12 hours.
- In 2017, Compass Languages invested in building a Neural Machine Translation (NMT) engine based upon the 4 million words of EN>JP translation already performed since 2014.
- In January 2018, the GSIC adopted this engine and a post-editing function to perform all alert translations.
- The move to Compass’ NMT/PE will save the company up to 40% of their annual translation cost.
This case study demonstrates Compass Languages commitment to meet customer needs, adapt to changing customer requirements, and innovating new processes and technologies to lower costs.