Best Practices

This article is a compilation of translation productivity tips.

How can I increase my productivity in the editor?

1. Don't focus on the suggestions. Typing the translation directly can often be faster than waiting for the suggestions to update. Consider the machine output during pauses in typing.

2. Ignore the machine suggestions if and when you want to. When you know the translation, and you're a fast typist, the machine assistance will only slow you down.

3. Post-edit if and when you want to. Press Shift+Enter to insert the full suggestion. This technique can be faster when you only want to make minor changes.

4. Learn the keyboard shortcuts; don't use the mouse.

5. Learn the copy-source-to-target hotkey. The MT system will corrupt non-linguistic strings composed of numbers, equations, and punctuation.

6. For CJK input (Chinese, Japanese, Korean), we recommend Google Input Tools for Chrome. We have not tested other CJK input extensions.

7. Configure your browser spell checker for the target language.

8. Try the Lexicon first for terminology. If you don't find what you need, then switch to another resource. Tab switching is slow. Try to avoid it. 

How can I improve Lilt's suggestions?

Lilt is an adaptive system, which means that it needs data (specifically, source/target pairs) for adaptation. Here are some rules of thumb for improving translation quality:

• Lilt works best when custom memories are created for each domain. So you might have memories for software, legal, medical, etc. A common mistake that we observe is aggregating all data into the default memories. This limits the system's ability to adapt to any one domain.
• 20k segments / memory seems to be the point at which people observe noticeable increases in translation quality. Uploading relevant TM data is the best way to improve suggestions for a domain.
• Google Translate / Microsoft Translator are admittedly broader domain systems since they are trained on more data. Baseline (i.e., unadapted) quality of Lilt is competitive for some domains, better for some domains (e.g., medical), and worse for some domains (e.g., software strings).
• Fine-grained lexical distinctions for common words (e.g., "Party") are the hardest for the system to learn. It has likely seen that word millions of times in the training data, and so it will need to see your preference many times to learn the distinction reliably. You should find that it learns rarer words and phrases faster. 

What is the most efficient way to review?

In the editor, confirmed segments are reduced in size so that more document context can fit on the screen. Clicking a confirmed segment will unconfirm it for editing. You must confirm the segment again to save it. When translating long documents, you may accidentally unconfirm or skip a segment. You can use the filters located in the white header bar to quickly locate untranslated/unsaved segments.

The QA Mode is also ideal for reviewing and making quick edits.