C2LIR: Continual Cross-Lingual Transfer for Low-Resource Information Retrieval

This paper proposes a method to train information retrieval (IR) model for a low-resource language with a small corpus and no parallel sentences. Although neural IR models based on pretrained language models (PLMs) have shown high performance in high-resource languages (HRLs), building PLM for LRLs is challenging. We propose (CLIR)-L-2, a method to build a high-performing neural IR model for LRL, with dictionary-based pretraining objectives for cross-lingual transfer from HRL. Experiments on the monolingual and cross-lingual IR in diverse low-resource scenarios show the effectiveness and data efficiency of (CLIR)-L-2.