KIT | KIT-Bibliothek | Impressum | Datenschutz

Answer-Convention Alignment for Agentic Text-to-SPARQL: AMA-KBQA at the WikiKGQA 2026 Challenge

Duan, Yannic; Klattig, Maximilian; Bardenhagen, Maurice; Li, Yuyang ORCID iD icon 1; Käfer, Tobias ORCID iD icon 1
1 Institut für Angewandte Informatik und Formale Beschreibungsverfahren (AIFB), Karlsruher Institut für Technologie (KIT)

Abstract:

We describe our submission to the WikiKGQA 2026 challenge, which asks systems to return exact answer sets for natural-language questions over Wikidata, scored by Macro QALD F1. Our entry adapts AMA-KBQA, a multi-agent KBQA framework with a graph- and model-agnostic reasoning core without any model training. Because QALD F1 rewards reproducing the benchmark’s modelling conventions rather than merely a semantically correct answer, the adaptation centres on answer-convention alignment. The agent explores the knowledge graph through typed tools that insert its arguments into valid SPARQL templates, commits a single query whose executed bindings form the scored answer, and passes that commitment through per-question convention rules and a deterministic repair chain that steer the answer toward the gold form. The convention rules themselves are hand-derived from the training gold; the reusable contribution is the mechanism around them: commit-time repair, answer-sanity gating, and self-consistency voting over executed answer sets. For the without-mentions tracks, where no entity or property identifiers are supplied, we add live entity linking via wbsearchentities and a hybrid retrieval pipeline over Wikidata’s properties. ... mehr


Zugehörige Institution(en) am KIT Institut für Angewandte Informatik und Formale Beschreibungsverfahren (AIFB)
Publikationstyp Proceedingsbeitrag
Publikationsjahr 2026
Sprache Englisch
Identifikator ISSN: 1613-0073
KITopen-ID: 1000199365
Erschienen in 1st Wiki-Based Knowledge Graph Question Answering Challenge, WikiKGQA 2026
Veranstaltung 1st Wiki-Based Knowledge Graph Question Answering Challenge (WikiKGQA 2026), Bari, Italien, 25.10.2062 – 29.10.2062
Verlag CEUR-WS
Serie 4284
Bemerkung zur Veröffentlichung in press
Externe Relationen Siehe auch
Schlagwörter KBQA; Large Language Models; Model Context Protocol; QALD; Text-to-SPARQL; Wikidata
Nachgewiesen in Scopus
KIT – Die Universität in der Helmholtz-Gemeinschaft
KITopen Landing Page