Replace roberta embedding with bge_base embedding in layoutlmv3

I’m experimenting Funsd like dataset for token classification using layoutLmv3, I’m trying to replace roberta embedding with bge_base, Is it possible to replace the embedding? will it improve the accuracy?