A Simple Geometric Method for Cross-Lingual Linguistic Transformations with Pre-trained Autoencoders

TL;DR

This paper introduces a simple geometric method to manipulate linguistic properties in multilingual sentence embeddings without retraining, validated across multiple languages and properties.

Contribution

It proposes a novel geometric mapping technique for cross-lingual linguistic transformations in pre-trained autoencoders, avoiding additional model tuning.

Findings

Effective manipulation of linguistic properties demonstrated

Works in both monolingual and cross-lingual settings

No additional training required for the autoencoder

Abstract

Powerful sentence encoders trained for multiple languages are on the rise. These systems are capable of embedding a wide range of linguistic properties into vector representations. While explicit probing tasks can be used to verify the presence of specific linguistic properties, it is unclear whether the vector representations can be manipulated to indirectly steer such properties. For efficient learning, we investigate the use of a geometric mapping in embedding space to transform linguistic properties, without any tuning of the pre-trained sentence encoder or decoder. We validate our approach on three linguistic properties using a pre-trained multilingual autoencoder and analyze the results in both monolingual and cross-lingual settings.