On Learning Language-Invariant Representations for Universal Machine Translation

Han Zhao, Junjie Hu, Andrej Risteski
Proceedings of the 37th International Conference on Machine Learning, PMLR 119:11352-11364, 2020.

Abstract

The goal of universal machine translation is to learn to translate between any pair of languages. Despite impressive empirical results and an increasing interest in massively multilingual models, theoretical analysis on translation errors made by such universal machine translation models is only nascent. In this paper, we formally prove certain impossibilities of this endeavour in general, as well as prove positive results in the presence of additional (but natural) structure of data. For the former, we derive a lower bound on the translation error in the many-to-many translation setting, which shows that any algorithm aiming to learn shared sentence representations among multiple language pairs has to make a large translation error on at least one of the translation tasks, if no assumption on the structure of the languages is made. For the latter, we show that if the paired documents in the corpus follow a natural \emph{encoder-decoder} generative process, we can expect a natural notion of “generalization”: a linear number of language pairs, rather than quadratic, suffices to learn a good representation. Our theory also explains what kinds of connection graphs between pairs of languages are better suited: ones with longer paths result in worse sample complexity. We believe our theoretical insights and implications contribute to the future algorithmic design of universal machine translation.

Cite this Paper


BibTeX
@InProceedings{pmlr-v119-zhao20b, title = {On Learning Language-Invariant Representations for Universal Machine Translation}, author = {Zhao, Han and Hu, Junjie and Risteski, Andrej}, booktitle = {Proceedings of the 37th International Conference on Machine Learning}, pages = {11352--11364}, year = {2020}, editor = {III, Hal Daumé and Singh, Aarti}, volume = {119}, series = {Proceedings of Machine Learning Research}, month = {13--18 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v119/zhao20b/zhao20b.pdf}, url = {https://proceedings.mlr.press/v119/zhao20b.html}, abstract = {The goal of universal machine translation is to learn to translate between any pair of languages. Despite impressive empirical results and an increasing interest in massively multilingual models, theoretical analysis on translation errors made by such universal machine translation models is only nascent. In this paper, we formally prove certain impossibilities of this endeavour in general, as well as prove positive results in the presence of additional (but natural) structure of data. For the former, we derive a lower bound on the translation error in the many-to-many translation setting, which shows that any algorithm aiming to learn shared sentence representations among multiple language pairs has to make a large translation error on at least one of the translation tasks, if no assumption on the structure of the languages is made. For the latter, we show that if the paired documents in the corpus follow a natural \emph{encoder-decoder} generative process, we can expect a natural notion of “generalization”: a linear number of language pairs, rather than quadratic, suffices to learn a good representation. Our theory also explains what kinds of connection graphs between pairs of languages are better suited: ones with longer paths result in worse sample complexity. We believe our theoretical insights and implications contribute to the future algorithmic design of universal machine translation.} }
Endnote
%0 Conference Paper %T On Learning Language-Invariant Representations for Universal Machine Translation %A Han Zhao %A Junjie Hu %A Andrej Risteski %B Proceedings of the 37th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2020 %E Hal Daumé III %E Aarti Singh %F pmlr-v119-zhao20b %I PMLR %P 11352--11364 %U https://proceedings.mlr.press/v119/zhao20b.html %V 119 %X The goal of universal machine translation is to learn to translate between any pair of languages. Despite impressive empirical results and an increasing interest in massively multilingual models, theoretical analysis on translation errors made by such universal machine translation models is only nascent. In this paper, we formally prove certain impossibilities of this endeavour in general, as well as prove positive results in the presence of additional (but natural) structure of data. For the former, we derive a lower bound on the translation error in the many-to-many translation setting, which shows that any algorithm aiming to learn shared sentence representations among multiple language pairs has to make a large translation error on at least one of the translation tasks, if no assumption on the structure of the languages is made. For the latter, we show that if the paired documents in the corpus follow a natural \emph{encoder-decoder} generative process, we can expect a natural notion of “generalization”: a linear number of language pairs, rather than quadratic, suffices to learn a good representation. Our theory also explains what kinds of connection graphs between pairs of languages are better suited: ones with longer paths result in worse sample complexity. We believe our theoretical insights and implications contribute to the future algorithmic design of universal machine translation.
APA
Zhao, H., Hu, J. & Risteski, A.. (2020). On Learning Language-Invariant Representations for Universal Machine Translation. Proceedings of the 37th International Conference on Machine Learning, in Proceedings of Machine Learning Research 119:11352-11364 Available from https://proceedings.mlr.press/v119/zhao20b.html.

Related Material