Multi-Hypothesis Distillation of Multilingual Neural Translation Models for Low-Resource Languages

Abstract

This paper explores knowledge distillation (KD) of multilingual pre-trained encoder-decoder and large language models for machine translation (MT). We argue that the teacher model's output distribution holds valuable insights for the student, beyond the approximated mode obtained through beam search (the standard decoding method for MT), and present Multi-Hypothesis Distillation (MHD), a sequence-level KD method that generates multiple translations for each source sentence. This provides a larger representation of the teacher model distribution and exposes the student model to a wider range of target-side prefixes. We explore n-best lists from beam search to guide the student's learning and examine alternative decoding methods to address issues like low variability and the under-representation of infrequent tokens. For low-resource languages, our research shows that while sampling methods may slightly compromise translation quality compared to beam search based approaches, they enhance the generated corpora with greater variability and lexical richness. This ultimately improves student model performance and mitigates the gender bias amplification often associated with KD. Our experiments show that MHD improves over existing KD techniques both when training encoder-decoder students from scratch and when distilling into pre-trained decoder-only models.