Modality Matching Matters: Calibrating Language Distances for Cross-Lingual Transfer in URIEL+

TL;DR

This paper tackles the limitations of URIEL+ by introducing modality‑matched language distances for cross‑lingual transfer. It develops three structure‑aware representations—geography as speaker distributions, genealogy in hyperbolic space, and typology as low‑noise latent islands—and unifies them with a composite, normalised distance. Across LangRank transfer benchmarks, these representations yield task‑dependent but generally improved transfer language selection, highlighting that no single metric dominates all tasks. The work also provides a general baseline composite distance and releases code and select resources to enable principled, reusable evaluation of linguistic distances. Overall, it offers a flexible toolkit to tailor language similarity signals to specific multilingual NLP applications.

Abstract

Existing linguistic knowledge bases such as URIEL+ provide valuable geographic, genetic and typological distances for cross-lingual transfer but suffer from two key limitations. One, their one-size-fits-all vector representations are ill-suited to the diverse structures of linguistic data, and two, they lack a principled method for aggregating these signals into a single, comprehensive score. In this paper, we address these gaps by introducing a framework for type-matched language distances. We propose novel, structure-aware representations for each distance type: speaker-weighted distributions for geography, hyperbolic embeddings for genealogy, and a latent variables model for typology. We unify these signals into a robust, task-agnostic composite distance. In selecting transfer languages, our representations and composite distances consistently improve performance across a wide range of NLP tasks, providing a more principled and effective toolkit for multilingual research.

Figures (1)

• Figure 1: A demonstration of URIEL+ language representations versus our proposed representations, for each modality. Distance scores are shown for URIEL+ (left number) and our proposed representation (right number). Lower values indicate greater similarity.