Constrained Density Matching and Modeling for Cross-lingual Alignment of Contextualized Representations

TL;DR

This paper introduces density-based alignment methods, Real-NVP and GAN-Real-NVP, for cross-lingual representation alignment that perform well with limited or no parallel data, improving multilingual NLP tasks.

Contribution

It proposes novel density matching and modeling techniques using Normalizing Flow for multilingual alignment, addressing data efficiency and training issues of prior methods.

Findings

Supervised approach with 20k sentences outperforms methods trained on 100k sentences.

Unsupervised approach enables removing parallel data without performance loss.

Validation criteria effectively guide training without relying on validation data.

Abstract

Multilingual representations pre-trained with monolingual data exhibit considerably unequal task performances across languages. Previous studies address this challenge with resource-intensive contextualized alignment, which assumes the availability of large parallel data, thereby leaving under-represented language communities behind. In this work, we attribute the data hungriness of previous alignment techniques to two limitations: (i) the inability to sufficiently leverage data and (ii) these techniques are not trained properly. To address these issues, we introduce supervised and unsupervised density-based approaches named Real-NVP and GAN-Real-NVP, driven by Normalizing Flow, to perform alignment, both dissecting the alignment of multilingual subspaces into density matching and density modeling. We complement these approaches with our validation criteria in order to guide the training process. Our experiments encompass 16 alignments, including our approaches, evaluated across 6 language pairs, synthetic data and 5 NLP tasks. We demonstrate the effectiveness of our approaches in the scenarios of limited and no parallel data. First, our supervised approach trained on 20k parallel data (sentences) mostly surpasses Joint-Align and InfoXLM trained on over 100k parallel sentences. Second, parallel data can be removed without sacrificing performance when integrating our unsupervised approach in our bootstrapping procedure, which is theoretically motivated to enforce equality of multilingual subspaces. Moreover, we demonstrate the advantages of validation criteria over validation data for guiding supervised training.