Wikontic: Constructing Wikidata-Aligned, Ontology-Aware Knowledge Graphs with Large Language Models

Abstract: Knowledge graphs (KGs) provide structured, verifiable grounding for LLMs, but current LLM-based systems commonly use KGs as auxiliary structures for text retrieval, leaving their intrinsic quality underexplored. In this work, we propose Wikontic, a multi-stage pipeline that constructs KGs from open-domain text by extracting candidate triplets with qualifiers, enforcing Wikidata-based type and relation constraints, and normalizing entities to reduce duplication. The resulting KGs are compact, ontology-consistent, and well-connected; on MuSiQue, the correct answer entity appears in 96% of generated triplets. On HotpotQA, our triplets-only setup achieves 76.0 F1, and on MuSiQue 59.8 F1, matching or surpassing several retrieval-augmented generation baselines that still require textual context. In addition, Wikontic attains state-of-the-art information-retention performance on the MINE-1 benchmark (86%), outperforming prior KG construction methods. Wikontic is also efficient at build time: KG construction uses less than 1,000 output tokens, about 3$\times$ fewer than AriGraph and $<$1/20 of GraphRAG. The proposed pipeline enhances the quality of the generated KG and offers a scalable solution for leveraging structured knowledge in LLMs.

• The paper introduces a system integrating LLM extraction with ontology alignment to construct knowledge graphs that adhere to Wikidata's schema.
• It employs iterative error correction and verification to reduce hallucinated triples and ensure factual consistency.
• Experiments show enhanced alignment accuracy and improved performance in entity-centric multi-hop question answering compared to unconstrained methods.

Wikontic: Constructing Wikidata-Aligned, Ontology-Aware Knowledge Graphs with LLMs

Overview

"Wikontic: Constructing Wikidata-Aligned, Ontology-Aware Knowledge Graphs with LLMs" (2512.00590) introduces a system for building knowledge graphs from unstructured text, utilizing LLMs for relation and entity extraction while rigorously aligning the output graphs with the formal ontological constraints and schema of Wikidata. The objective is to create KGs optimized for high-fidelity downstream reasoning and factuality by enforcing consistency with curated, collaborative resources, distinguishing the approach from purely open schema-free extractions.

Methodology

The Wikontic framework integrates state-of-the-art LLMs for entity and relation extraction and applies explicit, automated alignment steps to guarantee ontology compliance:

• Extraction: Leveraging LLMs (prompt-based and generative architectures) for initial candidate triples, synthesizing results from Seq2Seq extraction protocols [zhang2020minimize], REBEL-style language generation [cabot2021rebel], and neural relation extraction [distiawan2019neural, miwa2016end, zeng2014relation].
• Ontology Alignment: Every extracted triple undergoes mapping against the Wikidata ontology [vrande2012wikidata], using property types, entity classes, and cardinality constraints to validate admissibility. This enforces semantic coherence and eliminates structurally invalid tuples that could introduce spurious or misleading graph links.
• Error Correction and Verification: Incorporates iterative verification with ontology-based systems [chepurova2024prompt] to reduce hallucinated or mismatched relations, increasing ontological precision.
• Pipeline Efficiency: Optimizes prompt engineering [polat2025testing], retrieval-augmented generation [guo2024lightrag, han2024retrieval], graph memory [gutierrez2024hipporag, li2024graphreader], and dense IR methods [contriever] to scale the extraction pipeline and broaden knowledge coverage.

Key Results

The paper details several empirical findings:

• Alignment Accuracy: The Wikontic pipeline achieves substantially higher schema compliance than unconstrained LLM extractions, with fewer invalid triples due to misaligned property types or entity classes.
• Knowledge Coverage: Experiments demonstrate competitive fact recall benchmarked against open IE systems [stanovsky2018supervised, josifoski2021genie], but with improved precision stemming from ontological filtering. Coverage efficiency is comparable to other high-throughput KG synthesis frameworks [choubey2024distill].
• Downstream Utility: Ontology-aware KGs built with Wikontic facilitate superior performance in entity-centric multi-hop question answering (QA) as measured on tasks like HotpotQA [yang2018hotpotqa] and MuSiQue [trivedi2022musique], with fewer reasoning failures attributable to knowledge base inconsistency. They outperform hyper-relational graph approaches [panda2024holmes] on measures of answer factuality and type-correctness.
• Contradictory Claim: The paper finds that current LLMs, when combined with ontology-guided postprocessing, can construct KGs with reliability suitable for formal knowledge base population—contradicting prior assertions in the literature [wang2020language, mo2025kggen] that LLM-based extraction cannot robustly support ontological requirements without intensive human curation.

Theoretical and Practical Implications

The research demonstrates the viability of combining open information extraction (IE) with rigorous knowledge representation constraints to create scalable, trustworthy KGs for advanced reasoning. The ontological alignment protocol highlights the necessity of controlling schema drift in automatically generated KGs, suggesting an overview of linguistic and symbolic methods.

The results underscore the potential for integrating LLMs with structured KG resources to enhance trustworthiness in open-ended QA and other inference tasks [sui2024can]. This addresses the frequent critique that LLM-generated KGs are overly noisy or unreliable for formal reasoning.

From a systems perspective, Wikontic lays groundwork for further research into:

• Autonomous Curation: Adaptive, dynamic alignment protocols for continuous KG enrichment with minimal manual intervention.
• Schema Evolution: Automated methods for ontology extension and refinement based on large-scale extracted evidence.
• Hybrid Reasoning: Fusion of symbolic knowledge representation with sub-symbolic LLM inference for efficient multi-hop QA and entity reasoning.

Future Directions

Areas for future work include expansion to broader multilingual and multi-ontology extraction, integration with graph-based agent architectures [anokhin2024arigraph, li2024graphreader], and refinement of prompt and retrieval strategies for real-time KG updates. Potential investigations include leveraging KGs as non-parametric memory for continual LLM learning [gutierrez2025rag] and dynamically adapting ontologies in response to new discoveries.

Conclusion

Wikontic represents an authoritative advancement toward ontology-compliant, automatically constructed KGs from text using LLMs. By enforcing Wikidata-aligned structure and filtering, the system produces knowledge bases optimized for factuality, multi-hop inference, and downstream task reliability, demonstrating that automated knowledge graph construction need not sacrifice ontological rigor. This approach defines a clear path forward for both scalable KG synthesis and trustworthy, reasoning-ready AI systems.