sentence-transformers

/

distiluse-base-multilingual-cased

like 15

Also wondering what is the supported language list :)

Cheers, everyone
The language list is available at: "https://sbert.net/docs/sentence_transformer/pretrained_models.html." for each model card, under Training Data, there are specifications of the training data. In cases where the languages are omitted, you should check the training data.
For this model, the available languages on which the model can work are: "Arabic, Chinese, Dutch, English, French, German, Italian, Korean, Polish, Portuguese, Russian, Spanish, and Turkish."

Hello!

Thanks for your response! I have absolutely no idea how did I miss that :)

But you know what is interesting? I have already used it in my experiment (similarity search in Latvian and English dataset) and got a pretty solid recall result. But what is intriguing, the result is much better, almost twice, on long texts (multiple paragraphs) than on very short sentences (1 or 2 words).

Waaaaait, I just double checked and I think you are incorrect. The link, that you mentioned, I have already seen some time before and there is mentioned different model - distiluse-base-multilingual-cased-v1 While we are discussing the distiluse-base-multilingual-cased (without v1 on the end)

I believe Nils only had one dataset collection for multilingual models at the time, so I suspect that both models were trained on the same dataset. Having said that, the original distilbert may have been trained on more languages, and it's a common occurrence of multilingual models to have good performance on language X even if it was only pretrained on that language, but finetuned in others.

But what is intriguing, the result is much better, almost twice, on long texts (multiple paragraphs) than on very short sentences (1 or 2 words).

Perhaps multiple paragraphs is more similar to the data that was used during training than 1 or 2 words - I can imagine that would be the case.