When your Cousin has the Right Connections: Unsupervised Bilingual Lexicon Induction for Related Data-Imbalanced Languages

TL;DR

This work tackles unsupervised bilingual lexicon induction for extremely low-resource Indic languages that are closely related to a higher-resource language. It introduces a novel method that leverages a high-resource language MLM to translate masked words in the related low-resource language, iteratively growing a bilingual lexicon while replacing known words with their HRL equivalents to guide future predictions. Two orthographic rerankers—Basic (Levenshtein) and Rulebook (EM-learned substitution matrix)—allow effective selection of translation candidates, with multi-pass refinement improving coverage. Experiments on Bhojpuri and Magahi against Hindi, plus control languages Marathi and Nepali, show substantial improvements over strong baselines, and the authors release the resulting lexicons for five low-resource Indic languages. The approach is particularly suited to data-scarce settings with closely related HRLs, offering a practical resource for under-resourced language communities, while acknowledging limitations such as dependence on orthographic cognates and the need for related HRLs.

Abstract

Most existing approaches for unsupervised bilingual lexicon induction (BLI) depend on good quality static or contextual embeddings requiring large monolingual corpora for both languages. However, unsupervised BLI is most likely to be useful for low-resource languages (LRLs), where large datasets are not available. Often we are interested in building bilingual resources for LRLs against related high-resource languages (HRLs), resulting in severely imbalanced data settings for BLI. We first show that state-of-the-art BLI methods in the literature exhibit near-zero performance for severely data-imbalanced language pairs, indicating that these settings require more robust techniques. We then present a new method for unsupervised BLI between a related LRL and HRL that only requires inference on a masked language model of the HRL, and demonstrate its effectiveness on truly low-resource languages Bhojpuri and Magahi (with <5M monolingual tokens each), against Hindi. We further present experiments on (mid-resource) Marathi and Nepali to compare approach performances by resource range, and release our resulting lexicons for five low-resource Indic languages: Bhojpuri, Magahi, Awadhi, Braj, and Maithili, against Hindi.