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

TL;DR

This paper introduces Multi-Hypothesis Distillation (MHD), a sequence-level knowledge distillation method for multilingual translation models that generates multiple outputs to better capture the teacher's distribution, improving low-resource language translation and reducing bias.

Contribution

The paper proposes MHD, a novel sequence-level knowledge distillation approach that uses multiple hypotheses to enhance student model training for low-resource languages.

Findings

MHD improves translation quality for low-resource languages.

Sampling methods increase lexical richness and variability.

MHD reduces gender bias amplification in translation models.

Abstract

This paper explores sequence-level knowledge distillation (KD) of multilingual pre-trained encoder-decoder translation models. 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), 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 leverage $n$-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.