Machine translation post-editing (MTPE) is a process in which human translators review and improve the output generated by machine translation systems. Typically the output is generated by a Neural Machine Translation engine (NMT). In neural machine translation, the program’s neural network is responsible for encoding and decoding the source text, as opposed to running a set of predefined rules from the start. The quality of the output will depend on several factors, having to do with the source and target language, type of content to translate, etc.
Strengths:
- Well-structured, consistent, and unambiguous content (e.g. technical)
- Large volumes of repetitive content
- Low-stakes content
Weaknesses:
- Cultural factors and context
- Puns, metaphors, slogans
- Terminology, rare/unknown words
- Long sentences (depending on language)
- Lesser-known languages
This means that a document of product descriptions, for instance, is a good contender for MTPE, whereas a list of UI terms for software is not.
Machine Translation in MemoQ
It’s important to reiterate that memoQ is a CAT Tool, not a machine translation software. Instead, memoQ uses machine translation engines through standard MT plug-ins (e.g. DeepL, Google, etc.) In CAT tools, MT output can be intermingled with TM matches, and this is something to consider when selecting which output to use.
When working on an MTPE task, linguists should choose the most appropriate output from the Translation Results Pane, taking into account which match is more suitable for each specific segment. Please see an example of a Translation Results Pane below:
- Suggestions from translation memories or LiveDocs corpora are in Red. memoQ compares the current source segment to those stored in the different translation memories added to the project. The percentage indicates the proximity between the fuzzy match and the entry stored in the TM.
- Suggestions from term bases are Blue, whereas fragment search suggestions are Purple. These don’t always exist.
- In Orange we can finally see the MT suggestion. If you click that entry, you will see the provider’s symbol at the bottom, in this case it indicates that this output was generated by the Google Cloud Translation plugin.
Keep in mind that MT engines do not take TM entries into account (unless they are trained engines), and so their output will be the same no matter the amount of fuzzy matches that you have. In this specific example, the linguist would be able to choose between using one of the TM entries, or the MT-generated output. They have the added benefit of both fuzzy match and MT entry being visible in the translations pane, which could potentially allow the linguist to use parts from both suggestions in their own translation.
Tips and tricks for linguists working on MTPE
- Strive to use memoQ in a time-effective manner.
If you are inserting a suggestion from the Translation pane into the target segment for instance, you can simply double-click it to make it appear instantly as the target text; no need to manually copy and paste anything.
Learn the keyboard shortcuts or customize them as you please to open specific functions quicker: CTRL+K is the default shortcut to open Concordance; CTRL+G allows you to jump to a specific segment; F7 runs the spell-checker, etc. You can find a comprehensive list here: Keyboard shortcuts
- Find the optimal amount of editing.
Speed is one of the key advantages to MTPE, and if you find yourself altering MT-generated strings out of personal preference (instead of necessity), then you are not using machine translation to its full potential. - Prepare to edit machine translated and human translated texts differently.
When editing a text translated by a human translator, the errors you have to watch out for are mostly grammatical, such as conjugation or subject-verb agreement, and spotting typos. On the other hand, when working with a machine translated text, the errors tend to be mostly contextual: make sure that all segments are translated accurately within the context of the entire document. MT grammatical errors can also happen, especially in incorrect usage of adjectives or pronouns. Other common MT errors include lexical errors, such as inappropriate word choices or literal translations of idioms, and formatting errors, such as inconsistent spacing, punctuation, tag placement or capitalization. - There’s a difference between Light and Full MT Post-editing services.
Light MTPE involves correcting mistakes and reviewing some of the language choices, and making sure that the resulting text is comprehensible. It’s different from conventional (Full) MTPE, which requires the linguist to go deeper into editing the text for accuracy, clarity, flow, as well as local resonance. Codex currently works with Full MTPE only and expects it as standard service from all our post-editors, unless otherwise specified.
