Can the Variation of Model Weights be used as a Criterion for Self-Paced Multilingual NMT?

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Atrio, Àlex R., Allemann, Alexis, Dolamic, Ljiljana, Popescu-Belis, Andrei
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910634853007360
author Atrio, Àlex R.
Allemann, Alexis
Dolamic, Ljiljana
Popescu-Belis, Andrei
author_facet Atrio, Àlex R.
Allemann, Alexis
Dolamic, Ljiljana
Popescu-Belis, Andrei
contents Many-to-one neural machine translation systems improve over one-to-one systems when training data is scarce. In this paper, we design and test a novel algorithm for selecting the language of minibatches when training such systems. The algorithm changes the language of the minibatch when the weights of the model do not evolve significantly, as measured by the smoothed KL divergence between all layers of the Transformer network. This algorithm outperforms the use of alternating monolingual batches, but not the use of shuffled batches, in terms of translation quality (measured with BLEU and COMET) and convergence speed.
format Preprint
id arxiv_https___arxiv_org_abs_2410_04147
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Can the Variation of Model Weights be used as a Criterion for Self-Paced Multilingual NMT?
Atrio, Àlex R.
Allemann, Alexis
Dolamic, Ljiljana
Popescu-Belis, Andrei
Computation and Language
Many-to-one neural machine translation systems improve over one-to-one systems when training data is scarce. In this paper, we design and test a novel algorithm for selecting the language of minibatches when training such systems. The algorithm changes the language of the minibatch when the weights of the model do not evolve significantly, as measured by the smoothed KL divergence between all layers of the Transformer network. This algorithm outperforms the use of alternating monolingual batches, but not the use of shuffled batches, in terms of translation quality (measured with BLEU and COMET) and convergence speed.
title Can the Variation of Model Weights be used as a Criterion for Self-Paced Multilingual NMT?
topic Computation and Language
url https://arxiv.org/abs/2410.04147