Semantic Aware Linear Transfer by Recycling Pre-trained Language Models for Cross-lingual Transfer

  • 2025-05-22 12:54:30
  • Seungyoon Lee, Seongtae Hong, Hyeonseok Moon, Heuiseok Lim
  • 0

Abstract

Large Language Models (LLMs) increasingly incorporate multilingualcapabilities, fueling the demand to transfer them into target language-specificmodels. However, most approaches, which blend the source model's embedding byreplacing the source vocabulary with the target language-specific vocabulary,may constrain expressive capacity in the target language since the source modelis predominantly trained on English data. In this paper, we propose SemanticAware Linear Transfer (SALT), a novel cross-lingual transfer technique thatrecycles embeddings from target language Pre-trained Language Models (PLMs) totransmit the deep representational strengths of PLM-derived embedding to LLMs.SALT derives unique regression lines based on the similarity in the overlap ofthe source and target vocabularies, to handle each non-overlapping token'sembedding space. Our extensive experiments show that SALT significantlyoutperforms other transfer methods and achieves lower loss with acceleratingfaster convergence during language adaptation. Notably, SALT obtains remarkableperformance in cross-lingual understanding setups compared to other methods.Furthermore, we highlight the scalable use of PLMs to enhance the functionalityof contemporary LLMs by conducting experiments with varying architectures.

 

Quick Read (beta)

loading the full paper ...