Engines
- Greg Horváth (Unlicensed)
- Gábor Bessenyei (Globalese)
Engines are trained from corpora and are used to translate text. They are unidirectional, i.e. have one source language and one target language.
Types of engines
Custom
Domain-adapted, AI-boosted
Stock+
Stock
Custom engines
All engines trained in Globalese are custom engines. There is, however, a degree to their customization, depending on what combination of resources are used to create them.
Domain-adapted engines
Domain-adapted engines are engines containing both master and auxiliary/stock corpora, and trained using Globalese’s proprietary automated domain adaptation technology. By selecting your important in-domain TM(s) as “Master” training data, the engine will be focusing on the style and the terminology of those TM(s). You can choose to add generic stock data to extend the engine in case the volume of the in-domain data is not enough. You have also the option to add your own auxiliary data.
AI-Boosted engines
The option to train AI-boosted engines is available from Globalese V5. These engines are combining the terminology and style accuracy of a domain-adapted engine with the linguistic capabilities of large language models. In Globalese V5, AI-Boosted Engines use currently GPT models from OpenAI. The AI-boosting option is currently available for domain-adapted engines. AI-boosted engines typically provide the best results in most cases, especially as it is supporting dynamic keyword lists.
Please note that GPT models from MS Azure are currently available as a beta service. AI-boosted engines can only be used in the Cloud Text translation scenario.
Use case
The typical use case for domain-adapted engines is where adhering to a particular terminology and style is important. Some examples: product documentation, end-user manuals or software documentation, where it is essential to use the right terminology and style consistently.
Required training data
The following table shows the minimum and recommended number of segments.
Includes stock corpora? | Minimum volume (segments) | Recommended volume (segments) |
|---|---|---|
Yes | 15,000 master | 100,000+ master |
No | 15,000 master | 100,000+ master |
Typical training time
The typical training time for domain-adapted engines is between 10 and 28 hours.
Stock+ engines
Stock+ engines are customized stock engines, i.e. engines trained by extending a pre-trained stock engine with you own master data. The selected master data will be part of the engine. If there is new content in the master corpora, the engine will learn it. However, you should not expect changes in terminology and style preferences in the engine based on the master data added.
Use case
The typical use case for stock+ engines is where it is important to use a generic engine trained on a large data set, which is however incorporating your own training data too. You can also use this option if you the size of your own training data is not enough to train a domain-adapted engine. Some examples: annual reports, user generated content, web pages.
Required training data
A minimum of 1,000 segments, and a maximum of 1,000,000.
Typical training time
The typical training time for stock+ engines is between 10 minutes to 4 hours.
Stock engines
You can also use pre-trained stock engines for certain language combinations.