How to use with other languages other than english? about keybert

[MaartenGr]

Hmmm, it might improve if you increase the keyphrase_ngram_range to (1, 3) for example. However, this is exactly what can happen with KeyBERT. It is highly dependent on the underlying embedding model. For that reason, I typically like to use embedding models that are trained mostly for similarity measures, like those used in sentence-transformers (e.g., xlm-r-bert-base-nli-stsb-mean-tokens).

All in all, it requires some experimentation with models until you find the one best suited to your use-case.

[MaartenGr]

Yes! Simply use KeyBERT(model='xlm-r-bert-base-nli-stsb-mean-tokens') to use the multi-lingual model. There are several models that you could use here. However, the model that you referenced is the one I would suggest for any language other than English.

Also, you are in luck! A few days ago I released the v0.2 of KeyBERT which includes Flair. Flair allows you to, essentially, use any 🤗 transformers model. Thus, if you find any good French model, then you could also use that!

from keybert import KeyBERT
from flair.embeddings import TransformerDocumentEmbeddings

roberta = TransformerDocumentEmbeddings('roberta-base')
model = KeyBERT(model=roberta)

Then, instead of roberta-base, you can use any model from this list.

[LeMoussel]

Thank you for your help. I am a beginner with transformer model.
I found these 2 french models

• camembert-base : CamemBERT is a state-of-the-art language model for French based on the RoBERTa model.
• abhilash1910/french-roberta : This is a Masked Language Model trained with Roberta on a small French News Corpus(Leipzig corpora).

Do you think I can test with these 2 models?

[MaartenGr]

Although I have not tested these models myself, I do believe CamemBERT is used quite often for French texts. It should be no problem if you were to use those models. I am quite curious myself what the performance is, so let me know if you find out!

[LeMoussel]

If you are interested, I can share in this issue, my code comparing the results of keyword extraction with

• TF/IDF
• Yake
• KeyBert

[MaartenGr]

That would be great! I cannot read French well, so I would be dependent on your opinion regarding the quality of the keywords 😄

[LeMoussel]

Compare keyword extraction results, in French language, from TF/IDF, Yake, KeyBert: https://gist.github.com/LeMoussel/aa5f7c46c6cb09473e97b20eb2e13cc4
The quality of keywords with KeyBert seems "strange" to me. 🤔