Multilinguality as Sense Adaptation

TL;DR

This paper introduces SENSIA, a sense-level alignment method for multilingual models that improves cross-lingual transfer by explicitly aligning semantic representations, outperforming other methods on diverse language benchmarks.

Contribution

The paper proposes SENSIA, a novel sense-based alignment technique that enhances multilingual models by explicitly aligning semantic senses across languages, reducing data requirements.

Findings

SENSIA outperforms comparable multilingual alignment methods on diverse benchmarks.

SENSIA achieves competitive accuracy with significantly less target-language data.

Learned sense geometry remains consistent across languages and with English.

Abstract

We approach multilinguality as sense adaptation: aligning latent meaning representations across languages rather than relying solely on shared parameters and scale. In this paper, we introduce SENse-based Symmetric Interlingual Alignment (SENSIA), which adapts a Backpack language model from one language to another by explicitly aligning sense-level mixtures and contextual representations on parallel data, while jointly training a target-language language modeling loss to preserve fluency. Across benchmarks on four typologically diverse languages, SENSIA generally outperforms comparable multilingual alignment methods and achieves competitive accuracy against monolingual from-scratch baselines while using 2-4x less target-language data. Analyses of learned sense geometry indicate that local sense topology and global structure relative to English are largely preserved, and ablations show that the method is robust in terms of design and scale.