Managing AI Translate & TQI at global level
Note
If you use an XTM-provided LLM subscription, you must purchase an AI Words allowance for your entire XTM Cloud instance. AI Words are source words sent to the LLMs. Words from the translation memory (TM), terminology, the Style guide field, and context segments are NOT counted as AI Words. For details about AI Words calculation and deduction for projects, see AI Words in XTM Cloud. For pricing information, contact your account manager.
Important
To improve the accuracy of LLM-generated translations, especially if the necessary terminology and translation memory (TM) records either do not exist or are very limited, we strongly recommend that Project Managers use a Style Guide during project creation (for details, see Creating a project: Machine translation). As a user with an XTM Cloud Administrator role, you must create one or more Style Guides that can then be used by Project Managers. For details, see Style Guide tab.
The old Style Guide solution, created on the basis of the custom field (for details, see Data – Custom fields – Project custom fields), will be deprecated from XTM Cloud version 26.2.
As a user with the XTM Cloud Administrator role, you can make global-level AI Translate & TQI settings. These settings will be the system default. However, Project Managers can change these default settings at customer and project level.
Select Configuration > Settings > AI Control Center > Features.
In the Features tab screen, in the AI Translate & TQI section, select the Manage button.
Result: the AI Translate & TQI settings pane is displayed on the right-hand side of the Features tab screen.
In the AI Translate & TQI settings pane, make the required settings:
Set as default at Customer and Project level. Select the appropriate setting for this checkbox:
cleared (default) – by default, the AI Translate & TQI feature is deactivated for all existing and newly created customers and projects. If required, Project Managers can activate the feature at customer and project level.
selected – by default, the AI Translate & TQI feature is activated for all existing and newly created customers and projects. If required, Project Managers can deactivate the feature at customer and project level.
LLM subscription radio buttons. Select the appropriate option for this setting:
XTM-provided LLM – to use LLM licenses owned by XTM International.
Note
If you use an XTM-provided LLM subscription, you must purchase an AI Words allowance for your entire XTM Cloud instance. For more information, contact your account manager.
XTM uses ChatGPT and Gemini to run the AI Tranlste & TQI feature. To check which models are currently used, see the Supported models section in https://xtm.ai.transifex.com/docs#section/LLM-Configuration.
I bring my own key – to use one or more of your own LLM licenses. In this case, you must connect your XTM Cloud instance to at least one of the supported LLMs (for details, see LLM Configuration tab).
To check which Gemini models are currently supported, see the Supported models section in https://xtm.ai.transifex.com/docs#section/LLM-Configuration.
If you are using GPT through Azure, the Deployment names setting is displayed. Use it to select two deployment names. This is required even if you use two LLMs (GPT and Gemini) to run the feature. We recommend that you use one deployment name for a small model (for example, 4o Mini) and another for a large model (for example, 4o). If you do not do so, Project Managers will have to select them at customer level. Two deployment names are required to obtain the best quality results.
Note
If you are using your own LLM keys, the size of the source text, terminology, TM records, context segments, and the Style guide instructions can influence the LLM usage costs. For details, refer to your LLM subscription rules.
Confidence threshold section. Use these settings to set the minimum quality score for acceptable or good translations:
Unacceptable – set a threshold below which segments are considered poor quality and need correction or post-editing by a Linguist. To do so, enter a numeric value in the text field. Alternatively, use the slider displayed above the Unacceptable label: the corresponding numeric value is displayed automatically in the text field. The Unacceptable quality range is marked in red on the slider. Target segments within this range are also marked in red in XTM Workbench. The default value is 0%.
Acceptable – set a threshold to create a range for medium-quality segments. These LLM-translated segments meet basic standards, but a Linguist might need to make minor improvements to them. To set the threshold, enter a numeric value in the text field. Alternatively, use the slider displayed above the Acceptable label: the corresponding numeric value is displayed automatically in the text field. The Acceptable quality range lies between the Unacceptable threshold, or 0% (if the Unacceptable threshold has not been set), and the Acceptable threshold value. It is marked in yellow on the slider. Target segments within this range are also marked in yellow in XTM Workbench. The default value is 0%.
Good – the setting value is fixed at 100%. Segments in the range between the Acceptable threshold, Unacceptable threshold (if Acceptable has not been set), or 0% (if neither the Acceptable nor Unacceptable threshold has been set), and 100% are considered high quality. They do not need corrections or post-editing by a Linguist. The Good quality range is marked in green on the slider. Target segments within this range are also marked in green in XTM Workbench.
Important
To ensure the best possible AI-generated translation quality, we recommend that you select the 95–100% Good quality range.
Send text for translation and scoring radio buttons. Select the appropriate option for this setting:
Important
If you use an XTM-provided LLM subscription, the option that you select will also influence the moment when the AI Words are deducted from your AI Words allowance.
XTM Workbench (default) – the text in a segment is only sent to an LLM when a user selects the segment in XTM Workbench, in Edit mode. The translation is then provided for the selected segment and the two subsequent ones (if they meet the MT/LLM conditions set for the project and are displayed in edit mode). So, the AI Words are then deducted for each of these three segments that have been successfully translated and scored by the feature.
After analysis – the whole text is sent to an LLM during project analysis. When a user enters XTM Workbench, all segments already have translations and scores (if they meet the MT/LLM conditions set for the project and are displayed in edit mode). Depending on the project settings, these translations are either inserted in the target segment, or only displayed in the Matches docked panel. So, the AI Words are deducted during project analysis for all project segments that have been successfully translated and scored by feature.
Important
The After analysis setting is required if the Intelligent Workflow tool is also to be implemented in your XTM Cloud instance. For details, see Global Intelligent Workflow settings.
If the AI Translate & TQI feature is set to run in analysis and XTM Terminology is selected in your XTM Cloud instance, you, as a user with an Administrator role, must enable the Match terms in analysis setting for the entire XTM Cloud instance (for details, see Terminology engines). Alternatively, Project Managers must enable this setting at customer level (for details, see Customer Editor – Settings). Otherwise, the AI Translate & TQI feature will not have access to terminology.
After analysis and in XTM Workbench – the whole text is sent to an LLM during project analysis. The text can be resent to the LLM on a segment-to-segment basis as in the In XTM Workbench option. So, during project analysis, the AI Words are deducted for all project segments that have been successfully translated and scored by the AI Translate & TQI feature. However, if the LLM fails to translate and score some segments during project analysis, a user can trigger this action in XTM Workbench. To do so, they must select an untranslated segment in XTM Workbench in Edit mode. If LLM-translation and -scoring succeed at this manual attempt, the AI Words for this newly translated and scored segment will now be deducted. AI Words will NOT be deducted again for two subsequent segments if they have already been translated and scored during project analysis.
Use auto inline tag placement radio buttons. Currently, it is not possible to change this setting from its default, Never. It means that inline tags are NOT processed by the auto-insert mechanism. XTM Cloud retains the structure of the inline tags inside matches as received from the LLM.
In the bottom right-hand corner of the AI Translate & TQI settings pane, select:
Cancel to close the AI Translate & TQI settings pane without saving your changes.
Save to save your changes.
Activate if you are activating and configuring the AI Translate & TQI feature at the same time.
Result: you have made the relevant global-level AI Translate & TQI settings. If required, Project Managers can change these settings at customer and project level.
If required, you can deactivate AI Translate & TQI at global level. For details, see Deactivating AI Translate & TQI at global level.