Large multilingual language models typically share their parameters across all languages, which enables cross-lingual task transfer, but learning can also be hindered when training updates from different languages are in conflict. In this article, we propose novel methods for using language-specific subnetworks, which control cross-lingual parameter sharing, to reduce conflicts and increase positive transfer during fine-tuning. We introduce dynamic subnetworks, which are jointly updated with the model, and we combine our methods with meta-learning, an established, but complementary, technique for improving cross-lingual transfer. Finally, we provide extensive analyses of how each of our methods affects the models.
%0 Journal Article
%1 10.1162/coli_a_00482
%A Choenni, Rochelle
%A Garrette, Dan
%A Shutova, Ekaterina
%D 2023
%J Computational Linguistics
%K low-resource-nlp related_works
%P 1-29
%R 10.1162/coli_a_00482
%T Cross-Lingual Transfer with Language-Specific Subnetworks for Low-Resource Dependency Parsing
%U https://doi.org/10.1162/coli\_a\_00482
%X Large multilingual language models typically share their parameters across all languages, which enables cross-lingual task transfer, but learning can also be hindered when training updates from different languages are in conflict. In this article, we propose novel methods for using language-specific subnetworks, which control cross-lingual parameter sharing, to reduce conflicts and increase positive transfer during fine-tuning. We introduce dynamic subnetworks, which are jointly updated with the model, and we combine our methods with meta-learning, an established, but complementary, technique for improving cross-lingual transfer. Finally, we provide extensive analyses of how each of our methods affects the models.
@article{10.1162/coli_a_00482,
abstract = {{Large multilingual language models typically share their parameters across all languages, which enables cross-lingual task transfer, but learning can also be hindered when training updates from different languages are in conflict. In this article, we propose novel methods for using language-specific subnetworks, which control cross-lingual parameter sharing, to reduce conflicts and increase positive transfer during fine-tuning. We introduce dynamic subnetworks, which are jointly updated with the model, and we combine our methods with meta-learning, an established, but complementary, technique for improving cross-lingual transfer. Finally, we provide extensive analyses of how each of our methods affects the models.}},
added-at = {2023-11-09T09:32:42.000+0100},
author = {Choenni, Rochelle and Garrette, Dan and Shutova, Ekaterina},
biburl = {https://www.bibsonomy.org/bibtex/2fb4870a21e0ef8a4eb48182258bb9c68/tomvoelker},
doi = {10.1162/coli_a_00482},
eprint = {https://direct.mit.edu/coli/article-pdf/doi/10.1162/coli\_a\_00482/2141924/coli\_a\_00482.pdf},
interhash = {e9088d3de51abe1f9d5d59b36bbcde51},
intrahash = {fb4870a21e0ef8a4eb48182258bb9c68},
issn = {0891-2017},
journal = {Computational Linguistics},
keywords = {low-resource-nlp related_works},
month = {07},
pages = {1-29},
timestamp = {2023-11-09T09:32:42.000+0100},
title = {Cross-Lingual Transfer with Language-Specific Subnetworks for Low-Resource Dependency Parsing},
url = {https://doi.org/10.1162/coli\_a\_00482},
year = 2023
}