Microsoft Case Study

• Collaborative action with +300 linguists.
• Project management team of 4 tech-savvy people.
• Quality assurance team of 4 dedicated people.
• MT Knowledge.
• Domain Knowledge.
• Linguistic Knowledge.
• Practical experience with MT post-editing.
• Scientific knowledge on computer science.
• Academic background on machine translation and machine learning.
• Extensive knowledge and experience of managing remote working teams.
• Industry practice on various file types.

• Post-edit large volumes of content.
• Meet tight release schedules.
• Reduce localization costs.
• Manage different quality expectations.
• Work with +300 linguists.

Samples

Task Definitions and Guidelines

• Review process is driven by “comprehensibility”.
• Pre-process focuses on reducing edits required.
• Linguists focus on improving key information.
• Quality evaluation through practical LQA tasks.
• Terminology is important.
• Feedback and query loop is crucial.

Using these guidelines, our post-editor team learn to make only necessary changes to the MT output by;

• adding meaning that is missing from the MT output.
• removing machine generated errors.
• reviewing words, terms and phrases, and making amendments when the original placement obscures the meaning of the text.

We expect from our linguists for;

• being comprehensible (i.e. an end-user understands the content of the message).
• accurate (i.e. it communicates the same meaning as the source text).
• stylistically fine, though the style may not be as good as that achieved by a native-speaker human translator.

Therefore they;

• aim for grammatically, syntactically and semantically correct translation,
• ensure that key terminology is correctly translated and that untranslated terms belong to the client’s list of “Do Not Translate” terms,
• ensure that no information has been accidentally added or omitted,
• edit any offensive, inappropriate or culturally unacceptable content,
• use as much of the raw MT output as possible,
• apply basic rules regarding spelling, punctuation and hyphenation.
• ensure that formatting is correct.

Linguists are also able to check the status of their progress in order to manage their time more efficiently.

• Guidelines and instructions were created and centralized across all client’s global projects, shared with all localization team via Smartcat project page and updated 24/7.
• Before launching the project, source contents were preprocessed. Important technical changes were applied to improve the quality, consistency and “readability” of the contents.
• Important comments were added to Smartcat before the kick-off.
• The source texts were uploaded to Smartcat Ecosystem and subcategories were created for each of the target languages.
• Thanks to the search and filter options of Smartcat Marketplace, post-editors were selected and multiple invitations were sent with ease.
• Questions and comments of the vendors were monitored via Smartcat Chat and Comment to Document feature 24/7.
• Since the deadline was pretty tight for such a big-scale project, more than 20 linguists worked on the files for each language pair.
• Different linguists working for the same file started to communicate via Smartcat comments section to ensure quality.
• The project was completed right on time after the last step of quality assurance.

• The first stage of quality assurance is the quality of the machine translated output. To ensure this, we applied pre- and post- linguistic engineering processes to the workflow.
• Post-editor’s goal is to make edits on the MT output to ensure quality by checking the meaning, putting the tags in their appropriate places and correcting the machine generated errors.
• In addition, a number of automatic and manual quality control mechanisms are in place at Localex during the editing process to catch careless errors or alert QA Team to potential problems.
• QA Team also implements the spot-checks in the files daily, and provide feedback to post-editors accordingly.

Pre-processing

The source files have be to pre-processed by our developer to make them localization friendly.

Post-processing

After the localization process over, the target files have to be post-processed to make them work on the client’s environment again.

Problem: Source files (xlf) are structurally not localization-friendly.

• The segmentation problem and the complex tag structure.
• Extremely long segments.
• On average, there are more than 2500 segments that are longer than 500 characters per language.
• Content has complicated tag structure; i.e. inline tags, markup tags, and placeholders.

• Xliff is a standard format customized from .xml and that’s why different tools provided the same results. It’s also not so easy to modify segmentation as the output file might be corrupted.
• We offered .xml parsing method to extract the source content and this enabled us to modify the source content accordingly.
• Each of the source segments in the .xlf files are extracted using an .xml parser and saved with their unique trans-unit IDs. These ID lines are not included in the translation production step.
• We could perform some normalization tasks on the content. Unnecessary line breaks are removed, tags are hidden, and segments are re-segmentated. The original pattern was preserved including segmentation, line breaks, whitespace characters, tag contents, URLs and the placeholders within double brackets.

Detailed QA process requirement

Production phase was followed by QA 1, import back, and QA2 steps. Within first QA step, team mainly focused on linguistic checks and the custom placeholders which required custom checklist items. Afterwards the translated output was imported back to their intended locations in the original files.

• Efficient scalability: 3.4 Mn words, 6 languages, 10 days.
• High-quality translations.
• Translation productivity increase 30-40% (5-8K words/day).
• 91% linguist satisfaction with the project.
• Reduced project costs: 35-40% savings.
• Minimal client investment.
• High-tech workflow automation.
• Seamless communication.
• Ongoing improvement through feedback from 300+ linguists.
• New data to train machines.
• New integration processes and platforms.