Timezone: »
The goal of universal machine translation is to learn to translate between any pair of languages, given pairs of translated documents for \emph{some} of these 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 take one step towards better understanding of universal machine translation by first proving an impossibility theorem in the general case. In particular, we derive a lower bound on the translation error in the many-to-one 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. On the positive side, we show that if the documents follow a natural encoder-decoder generative process, then we can expect a natural notion of ``generalization'': a linear number of pairs, rather than quadratic, suffices. 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 in terms of the total number of documents per language pair needed. We believe our theoretical insights and implications contribute to the future algorithmic design of universal machine translation.
Author Information
Han Zhao (Carnegie Mellon University)
PhD@CMU
Junjie Hu (Carnegie Mellon University)
Andrej Risteski (CMU)
More from the Same Authors
-
2021 : The Effects of Invertibility on the Representational Complexity of Encoders in Variational Autoencoders »
Divyansh Pareek · Andrej Risteski -
2023 Poster: How Do Transformers Learn Topic Structure: Towards a Mechanistic Understanding »
Yuchen Li · Yuanzhi Li · Andrej Risteski -
2023 Poster: Neural Network Approximations of PDEs Beyond Linearity: A Representational Perspective »
Tanya Marwah · Zachary Lipton · Jianfeng Lu · Andrej Risteski -
2022 Workshop: Principles of Distribution Shift (PODS) »
Elan Rosenfeld · Saurabh Garg · Shibani Santurkar · Jamie Morgenstern · Hossein Mobahi · Zachary Lipton · Andrej Risteski -
2021 Poster: Representational aspects of depth and conditioning in normalizing flows »
Frederic Koehler · Viraj Mehta · Andrej Risteski -
2021 Spotlight: Representational aspects of depth and conditioning in normalizing flows »
Frederic Koehler · Viraj Mehta · Andrej Risteski -
2020 Poster: Empirical Study of the Benefits of Overparameterization in Learning Latent Variable Models »
Rares-Darius Buhai · Yoni Halpern · Yoon Kim · Andrej Risteski · David Sontag -
2019 Poster: On Learning Invariant Representations for Domain Adaptation »
Han Zhao · Remi Tachet des Combes · Kun Zhang · Geoff Gordon -
2019 Oral: On Learning Invariant Representations for Domain Adaptation »
Han Zhao · Remi Tachet des Combes · Kun Zhang · Geoff Gordon