# Semantic Drift in Multilingual Representations

## TL;DR

This paper introduces a methodology to analyze semantic relations and drift in multilingual representations, revealing that models trained on monolingual data can preserve linguistic relations and reflect phylogenetic structures.

## Contribution

It proposes a novel approach using representational similarity analysis to compare languages and detect semantic drift without relying on etymological data.

## Key findings

- Multilingual models preserve language relations based on semantic organization.
- Phylogenetic trees reconstructed from representations align with linguistic expert classifications.
- Semantic drift can be quantitatively measured across language families.

## Abstract

Multilingual representations have mostly been evaluated based on their performance on specific tasks. In this article, we look beyond engineering goals and analyze the relations between languages in computational representations. We introduce a methodology for comparing languages based on their organization of semantic concepts. We propose to conduct an adapted version of representational similarity analysis of a selected set of concepts in computational multilingual representations. Using this analysis method, we can reconstruct a phylogenetic tree that closely resembles those assumed by linguistic experts. These results indicate that multilingual distributional representations which are only trained on monolingual text and bilingual dictionaries preserve relations between languages without the need for any etymological information. In addition, we propose a measure to identify semantic drift between language families. We perform experiments on word-based and sentence-based multilingual models and provide both quantitative results and qualitative examples. Analyses of semantic drift in multilingual representations can serve two purposes: they can indicate unwanted characteristics of the computational models and they provide a quantitative means to study linguistic phenomena across languages. The code is available at https://github.com/beinborn/SemanticDrift.