What does it mean to be language-agnostic? Probing multilingual sentence encoders for typological properties

TL;DR

This paper investigates how multilingual sentence encoders encode typological linguistic properties across languages and layers, revealing differences linked to pretraining strategies.

Contribution

It introduces probing methods to analyze typological properties in multilingual encoders and compares how these properties are distributed across model layers.

Findings

Different encoders encode linguistic variation differently.

Pretraining strategies influence how typological information is captured.

Layer-wise analysis reveals distribution patterns of linguistic features.

Abstract

Multilingual sentence encoders have seen much success in cross-lingual model transfer for downstream NLP tasks. Yet, we know relatively little about the properties of individual languages or the general patterns of linguistic variation that they encode. We propose methods for probing sentence representations from state-of-the-art multilingual encoders (LASER, M-BERT, XLM and XLM-R) with respect to a range of typological properties pertaining to lexical, morphological and syntactic structure. In addition, we investigate how this information is distributed across all layers of the models. Our results show interesting differences in encoding linguistic variation associated with different pretraining strategies.