This paper investigates whether injecting semantic structural knowledge of low-resource or unfamiliar languages into Large Language Models (LLMs) enhances performance on downstream Text-to-SQL tasks. We evaluate our approach on Galician, a Romance low-resource language, and, to demonstrate its generality, also on Guarani, a (very) low-resource language of an entirely distinct linguistic profile. Our empirical results show that semantically-aware models consistently outperform baselines across all benchmark metrics.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。