HierSQL: A Hierarchical Schema-Aware Encoder–Decoder Framework for Complex Natural Language to SQL Translation
Journal of Intelligent Systems and Applied Data Science · 06.07.2026 · 4 (1)
DOI: 10.64680/jisads.v4i1.69 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Intelligent Systems and Applied Data Science, online erschienen am 06.07.2026. Verfasst von Ayesha Mafrah A M und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Converting natural language queries into executable SQL statements commonly termed NL2SQL remains a fundamental challenge in the intersection of natural language processing and database systems. While transformer-based models have considerably advanced this field, three persistent bottlenecks continue to limit real-world deployment: (i) poor generalisation across unseen database schemas, (ii) inability to handle complex nested sub-queries and multi-table joins, and (iii) fragile schema linking under ambiguous user phrasing. This paper proposes HierSQL, a Hierarchical Schema-Aware Encoder–Decoder framework that addresses these limitations through three synergistic components: a dual-channel schema encoder that fuses column-level and table-level representations via cross-attention, a graph-augmented SQL decoder that leverages foreign-key relationships encoded as a heterogeneous relational graph, and a curriculum-guided multi-dataset training strategy that jointly trains on four complementary benchmarks. We evaluate HierSQL on BIRD NeurIPS'23, KaggleDBQA, SEDE (Stack Exchange Data Explorer), and ATIS, achieving execution accuracy of 72.4%, 48.3%, 61.7%, and 86.2% respectively, outperforming all supervised fine-tuning baselines and matching or exceeding prompt-based LLM approaches at a fraction of the inference cost. Our ablation study confirms the additive benefit of each module.
Zitieren
APA
M, A. M. A., & Mallappa, S. (2026). HierSQL: A Hierarchical Schema-Aware Encoder–Decoder Framework for Complex Natural Language to SQL Translation. Journal of Intelligent Systems and Applied Data Science, 4(1). https://doi.org/10.64680/jisads.v4i1.69
Vancouver
M AMA, Mallappa S. HierSQL: A Hierarchical Schema-Aware Encoder–Decoder Framework for Complex Natural Language to SQL Translation. Journal of Intelligent Systems and Applied Data Science. 2026;4(1). doi:10.64680/jisads.v4i1.69
BibTeX
@article{M2026,
author = {M, Ayesha Mafrah A and Mallappa, Suresha},
title = {HierSQL: A Hierarchical Schema-Aware Encoder–Decoder Framework for Complex Natural Language to SQL Translation},
journal = {Journal of Intelligent Systems and Applied Data Science},
year = {2026},
volume = {4},
number = {1},
doi = {10.64680/jisads.v4i1.69}
}
EUVIMED