Post-edition is the task of improving a machine translation (MT) output. This service is part of a wider workflow that may involve the preparation of the input, the implementation of MT and the evaluation of the obtained text. It’s a complex process that involves technology know-how, artificial intelligence and linguistic knowledge1 in its various steps.
In order to obtain a better output after implementing the MT engine, post-editors will prepare the source text. This is because there are texts that are more suitable for MT than others. Pre-editing is the process of preparing the source text before MT to obtain a better MT raw output. The most common actions required in this step are the following:
- Manage terminology
- Apply style guides
- Shorten sentence length
- Reduce long noun phrases
At this stage, the MT engine translates the source text. The device can be integrated in a CAT tool, it can be a client’s engine or Google Translate, among other options. Depending on the project’s scope or requirements, a sample may be machine-translated to check the output. According to the results — and if needed — the project’s team makes adjustments in the source text or the engine.
Depending on the client’s requests and needs, the translated output can be delivered without post-editing at all (raw output), or with light post-editing or deep post-editing. Regardless of which process is applied, there are certain rules that determine the post-editing process. According to the Translation Automation User Society (TAUS), during the post-editing task, the post-editor should bear in mind these rules:
- Do not retranslate the text
- Decide changes quickly (“2-second rule”)
- Translate the whole text, unless some phrases are classified as untranslatable
- Correct incomprehensible sentences
- Delete inaccurate sentences if they are irrelevant and difficult to correct
- Focus on semantic and syntactic mistakes
- Don’t correct stylistic errors (their correction is subject to prior agreement)
- Don’t replace recurring terms with synonyms
Feedback and Evaluation
When developing an MT engine, the post-editor not only corrects the text, but also provides feedback to the engineers. Usually, the evaluation is made using standardized forms. This is a very important step that helps improve the MT device. The MT team retrains the engine based on the feedback provided (changing configurations, uploading new bilingual samples, for instance). With this step, the engine is “trained” so the quality of the MT output improves gradually.
The Zero Step
Like in any other localization project, there is a step that cannot be skipped. For a successful delivery, it’s important to have a prior agreement with clients about what they expect of the MT workflow. Specifically, what kind of post-editing process will be applied (none, light, or deep), style preferences, proper nouns treatment, date format and untranslatable phrases, among others, are details that need to be specified before the project starts. This kind of agreement is the foundation of any localization task.
1As we can see in the chart, the skills and the expertise of linguists play a key part in the MT’s workflow.
- Post-editing Highlights: What to Correct
- Linguist Profiling: What Makes an Ideal Candidate for Post-editor
- Lead Linguist Bibiana Cirera’s View on Machine Translation
- 4 Stages and 8 Rules for Successful Post-editing