Semantic Reliability-Aware Representation Learning for Low-Resource Clinical Cross-Lingual Retrieval

Cross-lingual retrieval learning expands access to clinical information across languages, but standard contrastive learning treats aligned pairs as equally informative. I investigate whether improvement in retrieval beyond ordinary contrastive adaptation comes from incorporating the semantic importance of the training evidence. English-Swahili is used to train a reliability-aware residual projection over frozen XLM-R representation and compare it with an otherwise matched InfoNCE projection. Recall at 1 (R@1) is improved by the reliability-aware objective from \(0.7490\) to \(0.8980\) following the same direction across five evaluated seeds. However, important boundaries are revealed from controlled tests. Individual objective components are not uniformly retrieval-optimal, as the advantages do not hold across multilingual backbones. Hausa, Amharic, Yoruba and Zulu frozen transfer varies by language and retrieval direction. A falsification audit further demonstrates that weak supervision is informative but sensitive to lexical definition. Results provide proof that reliability-aware supervision can improve clinical retrieval, even though the benefits are dependent on the transfer and representation setting.