My task deals with multi-language like (english and hindi). For that I need a common embedding to represent both languages.
I know there are methods for learning multilingual embedding like 'MUSE' ( https://github.com/facebookresearch/MUSE ), but this represents those two embeddings in a common vector space, obviously they are similar, but not the same.
So I wanted to know if there is any method or approach that can learn to represent both embedding in form of a single embedding that represents the both the language.
Any lead is strongly appreciated!!!
Hi Rajesh Sahu ,
Google's multilingual BERT is an example for a 100+ languages neural network that can produce embeddings out of the box. You can train classifiers on top of it, train it with data from one language and predict on data from other languages:
https://github.com/google-research/bert/blob/master/multilingual.md
If you want to use this (or similar multilingual models using different languages, or ptr-train your own model) then I would recomment to use this python library, because they already have most pre-trained models available:
https://github.com/huggingface/transformers
anyway, the bert-base, multilingual cased model should be a good starting point for you, as it contains Hindi and English!
I suggest you to follow: https://www.aclweb.org/anthology/P16-1190.pdf
https://ruder.io/cross-lingual-embeddings/
https://github.com/babylonhealth/fastText_multilingual
I had some good direction with ttps://github.com/google-research/bert/blob/master/multilingual.md.
- Badges
- Science topic
- Similar topics
- Psychology
- Mental Processes
- Learning