Re:infer now supports multilingual sources and datasets. This means that Re:infer's models can understand sources that contain multiple different supported languages, without actually having to translate them.
The languages currently supported within multilingual sources and datasets are: English, French, German, Spanish, Italian, Portuguese and Dutch (we'll be expanding this list over time!).
What this means in practice is that if users work and do business in several languages that are supported by Re:infer, they can train Re:infer on verbatims in those languages, rather than translating everything into a single language.
Important considerations when looking to use multilingual sources and datasets:
- If a dataset is multilingual, users will not be able to see translations of any verbatims (as provided for translated datasets), so they will need to be able to understand all of the languages in the dataset to effectively train their model
- Understanding multiple languages is a more complex machine learning problem than understanding a single language, so these datasets may experience a slight drop in performance compared to datasets in English
- Re:infer will only be able to understand language from one of the supported languages listed above. If there are other languages present in the dataset, tagging these verbatims with labels used on verbatims in supported languages will be confusing for Re:infer. It is better to label these as their own specific labels that capture the language as a label, but Re:infer will not be able to interpret the specifics of the unsupported language
How do you create multilingual sources and datasets?
For both data source and datasets, the language family is selected when they are created, and cannot be changed once they are.
Simply select 'multilingual' from the language family dropdown on the create source or create dataset modal (it's typically the last setting to select).
Please Note: Multilingual datasets can contain sources of any language family that Re:infer supports.
For more detail on creating a source in the UI, see here.
For more detail on creating a dataset, see here.