0
$\begingroup$

I have a doubt solving a test. The idea here is to demonstrate the NLP and Machine Translation abilities.

The Dataset is a multilingual, multi-context set of documents. The dataset is divided on context categories (Wikipedia, conference_papers, Amazon Reviews, etc.,) and on languages.

The objective is to create a document cartegorization classifier (in Python) for the different contexts of the documents. The classifier has to be done at context level, regardless of the language the documents are written in.

An important fact is that The dataset original has been modified and a document Never is repeated in 2 languages.

I have 2 ideas on mind to solve that:

  1. Train on all the data creating a multilingual classifier
  2. Doing language detection first and use monolingual models later.

What could be a reasonable approach to doing text classification for multiple languages?

$\endgroup$
0
$\begingroup$

This depends on whether you have enough training data for all languages. If yes, doing language ID and language-specific models might be a good choice, especially if there is a BERT-like model available for each language.

An alternative would be, do the language ID and than machine-translate the input into English and only train an English classifier. You can use e..g, high-quality pre-trained Marian models recently published by the University of Helsinki.

Otherwise, I would use pre-trained multilingual representations (probably XLM-R that is much better than Multilingual BERT) to get representation and train a single classifier for all languages. The multilingual representations seem to have even some zero-shot abilities, i.e., the classifiers seem to generalize even for languages that are no in the training data.

| improve this answer | |
$\endgroup$
  • $\begingroup$ Thanks. I think I will try with XLM-RoBERTa. I guess this models have the same limitations of Standard BERT, who has a maximum input length of 512 tokens. Am I Right? $\endgroup$ – Marc Aug 14 at 17:28
  • $\begingroup$ Yes it has. But XML-R has much larger vocabulary, so on average, the 512 tokens can cover longer text. $\endgroup$ – Jindřich Aug 14 at 17:49

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.