Exploring Representational Disparities Between Multilingual and Bilingual Translation Models

TL;DR

This paper investigates how multilingual translation models differ from bilingual ones in their internal representations, revealing that multilingual models tend to have less isotropic and more language-specific representations, which impacts their performance.

Contribution

It provides a geometric analysis of representation disparities between bilingual and multilingual models, highlighting the role of language-specific information in these differences.

Findings

Multilingual models have less isotropic representations than bilingual models.

Multilingual representations occupy fewer dimensions, indicating reduced capacity.

Language-specific information contributes significantly to anisotropy in multilingual models.

Abstract

Multilingual machine translation has proven immensely useful for both parameter efficiency and overall performance across many language pairs via complete multilingual parameter sharing. However, some language pairs in multilingual models can see worse performance than in bilingual models, especially in the one-to-many translation setting. Motivated by their empirical differences, we examine the geometric differences in representations from bilingual models versus those from one-to-many multilingual models. Specifically, we compute the isotropy of these representations using intrinsic dimensionality and IsoScore, in order to measure how the representations utilize the dimensions in their underlying vector space. Using the same evaluation data in both models, we find that for a given language pair, its multilingual model decoder representations are consistently less isotropic and occupy fewer dimensions than comparable bilingual model decoder representations. Additionally, we show that much of the anisotropy in multilingual decoder representations can be attributed to modeling language-specific information, therefore limiting remaining representational capacity.