$AgenticT^2S$: Robust Text-to-SPARQL via Agentic Collaborative Reasoning over Heterogeneous Knowledge Graphs

ACL ARR 2026 January Submission669 Authors

Keywords: Knowledge Graph Question Answering (KGQA), Text-to-SPARQL, Multi-Agent Systems, Heterogeneous Knowledge Graphs, Schema-Aware Reasoning, Query Decomposition, SPARQL Verification

Abstract: Question answering over heterogeneous knowledge graphs (KGQA) requires reasoning across diverse schemas, partial alignments, and distributed endpoints. Existing text-to-SPARQL methods either depend on domain-specific fine-tuning or assume a single graph, limiting transfer to low-resource domains and hindering cross-graph queries. We present *AgenticT*²S, a modular framework that decomposes a question into subgoals handled by agents for retrieval, SPARQL generation, and verification. An allocator selects each subgoal to one or more candidate graphs using weak-to-strong schema alignment, and reduces to the single endpoint in the standard single-KG setting. To improve reliability, a two-stage verifier filters structurally invalid and semantically underspecified queries via symbolic validation and counterfactual consistency checks. Across three Wikidata benchmarks and a circular economy multi-KG benchmark, *AgenticT*²S improves execution accuracy and triple-level *F*1 by +21.28 and +22.83 percentage points, respectively, on average over the strongest baseline (AutoGen). On the heterogeneous circular economy benchmark, it reduces average input tokens by 46.4% (875 vs. 1632 ATU). Overall, the results show that agentic decomposition, schema-aware routing, and explicit verification yield more reliable text-to-SPARQL for heterogeneous, multi-graph KGQA.

Paper Type: Long

Research Area: AI/LLM Agents

Research Area Keywords: Question Answering

Languages Studied: Generation

Submission Number: 669