Title: DeSQ: A Decomposition-Based Approach to SPARQL Query Generation

Abstract:

Current methodologies for Knowledge Base Question Answering (KBQA) generally adhere to one of two prevailing models. The first involves generating formal queries, a process often characterized by fragility and a lack of transparent explainability. The second relies on directly retrieving answers by exploring the knowledge base, which tends to be computationally expensive and susceptible to hallucinations. To harness the advantages of both strategies while addressing their inherent limitations, we present DeSQ (Decomposition-based SPARQL Query Generation), a framework that remains agnostic to specific knowledge bases and functions through a three-step process.

Initially, DeSQ breaks down intricate questions into Atomic Constraints (ACs), which align with the relational architecture of the underlying knowledge base. Subsequently, the system produces a dual-component structured output: (a) a mapping that links each AC to its corresponding SPARQL fragment, utilizing standardized placeholders for variables and URIs, and (b) a URIs Grounding block that details each placeholder. Finally, the framework integrates these individual fragments to construct a fully formed SPARQL query.

DeSQ outperforms existing state-of-the-art methods on four of the five primary benchmarks and exhibits greater resilience against lexical variations. In addition to achieving higher performance, the framework streamlines the evaluation process by removing the requirement for a live KB endpoint. Furthermore, its structured output facilitates detailed error analysis, thereby supporting more precise and effective interventions for system improvement.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC