A Comparison of Language Modeling and Translation as Multilingual Pretraining Objectives

Zihao Li, Shaoxiong Ji, Timothee Mickus, Vincent Segonne, Jörg Tiedemann

Forskningsoutput: Kapitel i bok/rapport/konferenshandlingKonferensbidragVetenskapligPeer review

Sammanfattning

Pretrained language models (PLMs) display impressive performances and have captured the attention of the NLP community. Establishing best practices in pretraining has, therefore, become a major focus of NLP research, especially since insights gained from monolingual English models may not necessarily apply to more complex multilingual models. One significant caveat of the current state of the art is that different works are rarely comparable: they often discuss different parameter counts, training data, and evaluation methodology. This paper proposes a comparison of multilingual pretraining objectives in a controlled methodological environment. We ensure that training data and model architectures are comparable, and discuss the downstream performances across 6 languages that we observe in probing and fine-tuning scenarios. We make two key observations: (1) the architecture dictates which pretraining objective is optimal; (2) multilingual translation is a very effective pretraining objective under the right conditions. We make our code, data, and model weights available at https://github.com/Helsinki-NLP/lm-vs-mt.
Originalspråkengelska
Titel på värdpublikationProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing
RedaktörerYaser Al-Onaizan, Mohit Bansal, Yun-Nung Chen
Antal sidor13
UtgivningsortKerrville
FörlagThe Association for Computational Linguistics
Utgivningsdatum1 nov. 2024
Sidor15882-15894
ISBN (elektroniskt)979-8-89176-164-3
DOI
StatusPublicerad - 1 nov. 2024
MoE-publikationstypA4 Artikel i en konferenspublikation
EvenemangConference on Empirical Methods in Natural Language Processing - Miami, Förenta Staterna (USA)
Varaktighet: 12 nov. 202416 nov. 2024

Vetenskapsgrenar

  • 6121 Språkvetenskaper
  • 113 Data- och informationsvetenskap

Citera det här