I am making a model that uses encoded articles (multiple sentences). I have found the Universal Sentence Encoder by Tensorflow, but it says it is only for English. Specifically, I am looking for an encoder for the Macedonian language. Can I use this encoder and if not is there a multilingual model that understands Macedonian?
$\begingroup$
$\endgroup$
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
2
This Universal Sentence Encoder that you link is trained specifically on English data, so it's going to work very poorly on any other language (to be clear, it's likely to produce garbage).
Unfortunately it's quite unlikely that you'll find a similar pre-trained model for Macedonian. You would have to train your own model from Macedonian data, and you need a really large amount. Btw that's the main reason why these pre-trained models are often trained on English only, since there's a lot of English text available. In case you want to try this, there is a Macedonian corpus as part of the Universal Dependencies project.
-
$\begingroup$ I am planning on using it for article political bias classification. I have downloaded a lot of articles (>1Gb). I am sorry if this is a stupid question but is it better to train it on those articles or the wikipedia/UDP corpus? $\endgroup$– DaniJun 15, 2021 at 15:46
-
$\begingroup$ @Dani it's not a stupid question at all, don't worry. There are two advantages using your articles in order to obtain embeddings: 1) your data is probably larger 2) more importantly, with this option the meaning of the words is represented specifically for your data, whereas using another dataset might cause some inaccuracies. I'd suggest starting with traditional ML methods which don't require embeddings: it's simpler and faster, and later it can be used as baseline when you develop a more advanced model. $\endgroup$– ErwanJun 15, 2021 at 16:37