Improving Low-Resource Retrieval Effectiveness using Zero-Shot Linguistic Similarity Transfer

TL;DR

The paper tackles retrieval robustness when queries and documents come from related but distinct language varieties, a scenario common for low-resource and endangered languages. It introduces Zero-Shot Linguistic Similarity Transfer, a training method that fine-tunes neural rankers on translated query–document pairs from similar varieties and then transfers the learned patterns to unseen languages. Through experiments on mMARCO/v2, neuMARCO, and neuCLIR, it shows that pairwise language-variety fine-tuning improves performance on trained pairs and often transfers to unseen varieties, with mixed results across language families. The work highlights the potential for sustainable, resource-efficient cross-language IR and provides translated datasets and model checkpoints to advance future research. Overall, it suggests a paradigm shift toward leveraging linguistic similarity to extend neural IR capabilities to neglected languages, while acknowledging the need for further study on cross-family transfer.

Abstract

Globalisation and colonisation have led the vast majority of the world to use only a fraction of languages, such as English and French, to communicate, excluding many others. This has severely affected the survivability of many now-deemed vulnerable or endangered languages, such as Occitan and Sicilian. These languages often share some characteristics, such as elements of their grammar and lexicon, with other high-resource languages, e.g. French or Italian. They can be clustered into groups of language varieties with various degrees of mutual intelligibility. Current search systems are not usually trained on many of these low-resource varieties, leading search users to express their needs in a high-resource language instead. This problem is further complicated when most information content is expressed in a high-resource language, inhibiting even more retrieval in low-resource languages. We show that current search systems are not robust across language varieties, severely affecting retrieval effectiveness. Therefore, it would be desirable for these systems to leverage the capabilities of neural models to bridge the differences between these varieties. This can allow users to express their needs in their low-resource variety and retrieve the most relevant documents in a high-resource one. To address this, we propose fine-tuning neural rankers on pairs of language varieties, thereby exposing them to their linguistic similarities. We find that this approach improves the performance of the varieties upon which the models were directly trained, thereby regularising these models to generalise and perform better even on unseen language variety pairs. We also explore whether this approach can transfer across language families and observe mixed results that open doors for future research.

Figures (5)

• Figure 1: Our Zero-Shot Linguistic Similarity Transfer method leverages a neural ranker fine-tuned on a pair of language varieties, enabling robust retrieval across language varieties even in a zero-shot setting.
• Figure 2: Retrieval performance of the mMARCO original and translated queries on the documents of the original language variety. All models show significant drops in performance across language varieties.
• Figure 3: Performance comparison of baselines (Orig.) and our fine-tuned models (Ours) on the mMARCO French and Catalan queries on the French documents. mT5 recall scores were omitted as they are the same as BGE-M3.
• Figure 4: MRR@10 of baselines (Orig.) and our models (Ours) on the unseen mMARCO pairs. "mT5" shows the BGE-M3 results re-ranked with mT5.
• Figure 5: Recall@1000 of baselines (Orig.) and our models (Ours) on the unseen mMARCO pairs. mT5 scores were omitted as they are the same as BGE-M3.