.png){kind=icon}
.png){kind=icon}
Return
PPR-SQL: a robust text-to-SQL system via probabilistic graph pruning and multi-view reasoning
.png){kind=icon}
PRE
AI
DOI：10.1007/s10844-026-01049-4.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Deploying Text-to-SQL parsers in large-scale database environments faces two critical bottlenecks: structural disconnectivity from schema noise and reasoning fragility across varying query complexities. Existing retrieval-based methods often treat schemas as unstructured text, overlooking topological dependencies defined by foreign keys and thus causing frequent "broken join" failures. To bridge semantic retrieval and structural awareness, we propose PPR-SQL, a robust coarse-to-fine framework. At its core is a tuning-free Probabilistic Schema Pruning mechanism based on Personalized PageRank (PPR). By propagating relevance scores along foreign-key networks, it effectively recovers structurally essential bridge tables, achieving a 95.11% Strict Recall Rate (SRR) - substantially surpassing traditional semantic-only retrieval baselines while significantly compressing the schema search space. Complementing this structural grounding, Multi-View Reasoning synthesizes diverse SQL candidates from concise to schema-aware Chain-of-Thought perspectives, augmented by Active Value Grounding (AVG) to resolve entity-value mismatches and Intent-Consistency Reranking (ICR) for reliable selection. Extensive experiments on the BIRD development set demonstrate that PPR-SQL attains 70.34% execution accuracy (using Qwen3 + XiYan ensemble), markedly outperforming representative open-source baselines such as MAC-SQL (59.39%), RSL-SQL (67.21%), and single XiYan variants, with particular gains on challenging multi-hop queries (60.69% vs. 40-54% in prior methods), highlighting its effectiveness and deployability in real-world database systems.
Keywords:
Text-to-SQL
Large language models
Personalized pagerank
Schema linking
Multi-view reasoning
Journal
J
IF:
3.4
Papers:
75
Citations:
0