Cross-lingual transfer by forced alignment of embeddings in different languages

ideanlp

• start with pairs of document in 2 different languages (e.g. books)
• force the embeddings of the mode for the target language to be aligned with that of the model for the source language
  • using extra encoders to aligned the encoded embeddings ?
    • document embeddings must be calculated using the token embeddings
  • using a reward model ?