Historically, one of the key problems with machine translation was that the machine consistently made the same mistakes. This meant that post-editors had to make the same correction over and over. Asia Online is unique in the Statistical Machine Translation (SMT) market with its rich set of post-editing and post editing tools that contribute to a complete human feedback loop for continued translation quality improvement. These tools enable users to make corrections to the translated output, with the corrections being analyzed and leveraged by the Asia Online translation platform for greater quality on future translations.

The Post-Editing tools provide a rich environment that allows reviewing translators/post-editors to do any or all of the following:

1. Identify the most common kinds of errors e.g. unknown words
2. Get direct and immediate access to any web based dictionary and glossary resources
3. Make edit changes to the SMT system output to correct high frequency translation errors
4. Chat in real time on specific translations to enable basic collaboration on the cleanup process
5. Blog about the process or the material being worked on to again facilitate communication and collaboration on the translation task
6. Compare translations by different editors to determine what is the best approach
7. Track different versions of translations for the same segment so that the system can learn the human variances that occur with language
8. Track different opinions on the best translation as there is very rarely complete agreement on the “best” translation

The changes made in the Asia Online Post-Editing tool are directly fed back into the translation platform. This ensures that once errors are corrected, the corrected data can be used to permanently make modifications to the statics that determine the best translation quality and in doing so eliminate any future occurrence of the same error.

This feature allows Asia Online’s Enterprise Translation Platform to improve each time users provide feedback and over time raise both the base output quality and the overall productivity of translators and post-editors

Although raw un-edited machine translation can be useful, Asia Online understands that machine translation alone will not produce high enough quality for many kinds of translation tasks. Our tools have been designed as an integrated post-editing and human colaboration environment that enables translation professionals to focus on the finer details of translation.

Unlike other SMT or Rules Based Machine Translation (RBMT) systems that have a very slow and painful error correction process, the Asia Online approach allows the user to take immediate control of how the initial translation engine will improve and evolve. The greater the post-editing and post-editing efforts made into the initial engine the more quickly it will improve.

Asia Online offers a number of models to use out translation platform. Our Enterprise Translation Portal is available as a Software as a Service (SaaS) offering, and is designed to meet the needs of small to mid-sized organizations. Because Asia Online’s Enterprise Translation Portla is a SaaS solution, a large server farm provides the mass of computing resources behind each key process. This enables the platform to respond to key user requirements much more rapidly and easily than traditional machine translation solutions.

No automated translation engine can be expected to produce human quality translations out-of-the-box. We see that the process of evolution is a very deliberate and structured process that takes a predictable path. Asia Online provides a variety of tools and processes to facilitate the enhancement of a customer translation system. In the typical development of a translation engine at Asia Online the following process is observed and is common to most projects.

Asia Online is able to achievement rapid quality improvement because, unlike our competitors dirty data approach, we only use clean data to train our engines. Clean data means that when an error occurs, the error is usually due to gaps in knowledge coverage in the data. Errors caused in SMT systems built on dirty data are usually due to bad data, which is very difficult and costly to remove. Asia Online provides the tools to understand which data was used to determine a translation, which in turn allows gaps to be seen when they occur. These gaps can quickly be filled, ensuring that future translations do not have the same errors. We have a through process of machine and human checks to ensure only clean data is used in our engines.

The above graphic illustrates how rapidly the raw output of the automated translation system can improve. As a focused error identification and correction strategy is executed, the system can improve in a matter of weeks. The green indicates phrases that are ready to publish directly from the translation platform.

• Initial translation engine is developed using customer TM and other bilingual assets. This data is often blended with data from Asia Online baseline reserves to augment and supplement sparse customer data.
• The initial engine is tested with new material and the errors are analyzed by competent linguists to determine an initial clean-up strategy. As these systems are often to be used with largely predictable types of content it is possible to make adjustments to improve the system dramatically just by focusing on 10% of the phrases that occur most frequently in the body of the new translations. As these core phrases are verified, corrected and re-inserted into the initial corpus the system will improve in quality by all objective and subjective measurements.
• The slightly improved or corrected engine is then opened up to scrutiny by a larger group of core translators who will now run a larger set of new material through the system and again analyze and correct the errors. Any error that is corrected will not appear again and thus the system rises rapidly in quality. It is possible to run several sets of new and changing material to build further robustness into the system and increase its ability to handle similar types of new material. This is dependent on the customer requirements and expectations and Asia Online makes this process easy by providing a variety of tools to monitor and track this refinement process.



Find & Correct Spelling Errors and Add Translations for Unknown Words.

• Finally, the system can be released to its target users who also have the ability to also suggest changes to incorrect translations. These translations are sent to administrators who can setup processes to accept or reject suggested changes. However, as the body of changes builds up over time, the translation system gets better. Thus, the more the system is used and corrected the better it is likely to get. The combination of an active feedback and correction mechanism and use of the community to identify translation errors allows an Asia Online system to improve at a rate that has never been seen before in machine translation. This process generates masses of new TM and also continually improves the new output produced by the system.

Since an Asia Online system is a living and constantly changing system there are always multiple concurrent ways that the system is improving. Some of this is because of a growing corpus of clean data underlying the system, but also because we are constantly improving our spell checking, parsing, segmentation and alignment tools and technologies.

There are three ways that the basic translation engines will continue to improve over time and with use:

• New data from Asia Online publisher partners and general data gathering efforts to expand the number of domains that we cover, add vocabulary in specialized fields, add vernacular and colloquialisms which all boost baseline quality.
• Corrective Feedback from linguistic experts and translators within your localization team who can identify the major error patterns and dramatically enhance the overall quality by cleaning these errors out of the system. Typically, by focusing on 20% of the most used phrases and language it is possible to make a large positive impact on the bulk of the content.
• For very large bodies of content that simply cannot be done by an internal team in any meaningful time frame, Asia Online can also support a broad web based collaboration environment. Corrective Feedback from crowdsourcing (users who visit the site and make suggestions to improve translations) can also be handled and we have processes and procedures in place to manage this and control quality as shown below.