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

2025-10-21

AI Summary

Existing linguistic knowledge bases (e.g., URIEL+) provide geographic, genealogical, and typological distances with two key limitations: (1) they rely on homogeneous vector representations, failing to capture heterogeneous structural relationships among languages; and (2) they lack interpretable, task-agnostic mechanisms for distance aggregation. This work proposes the *Typologically Aligned Linguistic Distance Framework*, the first to jointly integrate structure-aware modeling with principled fusion. Specifically, it employs speaker-weighted distributions for geographic distance, hyperbolic space embeddings for genealogical distance, and latent variable models for typological distance—each tailored to its respective relational structure. A unified, robust composite distance metric is then derived. Evaluated on multilingual NLP tasks, the framework significantly improves cross-lingual transfer performance, demonstrating both the efficacy of structurally adaptive language representations and the generalizability of interpretable, compositional distance aggregation.

Technology Category

• Natural Language Processing: Machine Translation, Multilinguality, Cross-Lingual NLP
• Knowledge Representation and Reasoning: Knowledge Representation Languages
• Machine Learning: Structured Learning

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Semantics and Knowledge: Scalable techniques for the creation, curation, publication, maintenance, and consumption of large, Web-based, structured, reusable, knowledge graphs and ontologies
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

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.

Problem

• Improving cross-lingual transfer by calibrating language distances
• Creating structure-aware representations for linguistic data types
• Developing unified composite distances for multilingual NLP tasks

Innovation

• Structure-aware representations for diverse linguistic data types
• Hyperbolic embeddings model genealogical language relationships
• Composite distance unifies signals for cross-lingual transfer