MT Optimize

Content is machine translated from English by Phrase Language AI.

MT Optimize is a Phrase-developed generative AI capability.

MT Optimize uses Generative AI to improve and transform a translated document. It restores fluency and consistency, improves grammar and ensures that brand terminology is used accurately. Customers can transform target text to meet their brand voice and audience requirements by specifying the formality level to apply to a document.

Powered by an interaction with OpenAI’s latest models, MT Optimize retrieves source and target text, the project term base, formality requirements and content requirements and transforms content at speed. When a term base is assigned to the project, MT Optimize respects preferred and forbidden terms defined in the term base to guide the generated output. Other term base information is not currently used by MT Optimize.

MT Optimize is available as an action in Orchestrator and is configurable via the Orchestrator workflow editor. MT Optimize with the Human Review workflow template provides an example of one possible use case.

The MT Optimize action consumes AI Units (AIUs). Customers do not need their own OpenAI credentials to start using it.

Due to very large file sizes, .IDML (Adobe InDesign) files are not supported.

Use cases

Automating post-editing

MT Optimize refines translations to produce fluent, consistent text that adheres to brand terminology and uses the right register for a given audience, across every segment of a document. It works across multiple segments at once, using context to ensure consistency. MT Optimize allows customers to reduce post-editing cost by boosting the translation quality of machine translation (MT).

Exclude locked segments

When configuring MT Optimize in an Orchestrator workflow, users can control whether locked segments are included in AI processing.

In the Parameters section of the MT Optimize action, set Include locked segments as AI context to:

Yes (default)

Locked segments are included as context and consume AI Units (AIUs). Including locked segments provides more context and can improve output quality.
No

Locked segments are excluded and do not consume AIUs. Excluding locked segments may lower output quality.

This option is useful for jobs with a high proportion of locked segments that do not need to be considered during optimization.

Existing workflows must be republished for the change to take effect.

Interaction with translation quality metrics

Phrase Quality Performance Score (QPS)

Phrase QPS is an effective but not perfect measure of the impact of MT Optimize. It captures the overall improvement in translation quality that MT Optimize provides, with the following limitations:
- MT Optimize restores consistency between segments, while Phrase QPS evaluates translations at the individual segment level.
- MT Optimize uses a term base to fix errors in terminology adherence. Phrase QPS is unaware of terminology.
- MT Optimize can adjust formality level to meet use case specifications, whereas Phrase QPS is unaware of specific style needs, if they are not encoded in the source text itself.
Reference-based metrics

Reference-based metrics such as BLEU, COMET and ChrF3 can evaluate the output of MT Optimize to a certain extent, by comparing it to one or more reference translations to capture lexical and semantic improvements.

These metrics can detect the segment-level linguistic changes and improvements in accuracy made by MT Optimize, but only when the optimized translation more closely matches the reference. Where the reference does not reflect document-level consistency, terminology adherence or formality level, BLEU, COMET and ChrF will not pick up on the optimization of these factors.

Language support

Although MT Optimize may support many language combinations, we have determined a set of recommended language pairs selected to adhere to Phrase’s high-quality standards:

English
French
Italian
Spanish
Portuguese
German
Dutch
Swedish
Danish
Norwegian Bokmål
Norwegian Nynorsk
Finnish
Hungarian
Estonian
Czech
Russian
Polish
Bulgarian
Greek
Latvian
Chinese Simplified
Chinese Traditional
Japanese
Korean
Thai
Vietnamese
Indonesian
Arabic
Turkish
Hindi

MT Optimize in the CAT web editor

When opening a job that has been processed by MT Optimize, the CAT web editor displays the MT optimized notification toast.

MT Optimize post-editing can result in the following outputs:

Edited target

The segment text was modified by MT Optimize. This is indicated by the MT Optimize indication and the MT optimized tooltip in the score/origin column of the CAT pane.
Estimated target

The segment quality was estimated by MT Optimize and the new score is presented in the score/origin column of the CAT pane with the MT optimized indication.