DIAL-KG: Schema-Free Incremental Knowledge Graph Construction via Dynamic Schema Induction and Evolution-Intent Assessment

Knowledge Graphs (KGs) are foundational to applications such as search, question answering, and recommendation. Conventional knowledge graph construction methods are predominantly static, rely ing on a single-step construction from a fixed corpus with a prede f ined schema. However, such methods are suboptimal for real-world sce narios where data arrives dynamically, as incorporating new informa tion requires complete and computationally expensive graph reconstruc tions. Furthermore, predefined schemas hinder the flexibility of knowl edge graph construction. To address these limitations, we introduce DIAL KG, a closed-loop framework for incremental KG construction orches trated by a Meta-Knowledge Base (MKB). The framework oper ates in a three-stage cycle: (i) Dual-Track Extraction, which ensures knowledge completeness by defaulting to triple generation and switching to event extraction for complex knowledge; (ii) Governance Adjudica tion, which ensures the fidelity and currency of extracted facts to prevent hallucinations and knowledge staleness; and (iii) Schema Evolution, in which new schemas are induced from validated knowledge to guide subsequent construction cycles, and knowledge from the current round is incrementally applied to the existing KG. Extensive experiments demon strate that our framework achieves state-of-the-art (SOTA) performance in the quality of both the constructed graph and the induced schemas.

💡 Research Summary

**
The paper “DIAL‑KG: Schema‑Free Incremental Knowledge Graph Construction via Dynamic Schema Induction and Evolution‑Intent Assessment” addresses the fundamental limitations of traditional knowledge‑graph (KG) construction pipelines, which are typically static, rely on a fixed corpus, and depend on pre‑defined schemas. Such pipelines are ill‑suited for real‑world scenarios where data streams in continuously and the underlying knowledge evolves over time. To overcome these challenges, the authors propose DIAL‑KG, a closed‑loop framework that incrementally builds and evolves a KG without requiring a fixed ontology.

At the heart of DIAL‑KG lies a Meta‑Knowledge Base (MKB). The MKB stores three kinds of meta‑information: (1) Entity profiles that normalize names, aliases, and types; (2) Relation schemas that define static triple structures; and (3) Event schemas that capture dynamic, time‑sensitive facts as trigger‑role‑time tuples. The MKB is continuously updated and serves both as a governance hub and as a contextual memory for subsequent processing batches.

The framework operates in a three‑stage iterative loop for each incoming batch Bₖ:

1. Dual‑Track Extraction – The system first decides whether a sentence expresses a simple, stable fact or a complex, temporally‑rich event. Simple facts are extracted as triples (static track), while complex statements are represented as events (event track). In the cold‑start phase, extraction relies on few‑shot prompting of a large language model (LLM). Once the MKB contains schema proposals, the system retrieves the top‑K most relevant schemas (K=30) and injects them into the prompt, ensuring that newly generated facts respect the evolving schema while staying within the LLM’s context window. Entity and event mentions are normalized through intra‑batch clustering (embedding similarity for entities, role‑trigger similarity for events) followed by cross‑batch alignment with existing MKB profiles.

2. Governance Adjudication – Extracted knowledge passes three verification modules:

   • Evidence Verification checks that each fact is backed by textual evidence, mitigating hallucinations.
   • Logical Verification validates consistency against the current MKB schemas, preventing contradictory triples.
   • Evolutionary‑Intent Verification assesses whether an existing fact should be deprecated, based on temporal cues or contradictory new evidence. Deprecation is “soft”: the fact’s status is set to Deprecated rather than being physically removed, preserving a full provenance trail.

3. Schema Evolution – Verified facts are used to induce new relation and event schemas. For relations, the system clusters entities, infers type constraints via LLM prompting, and formulates domain/range specifications. For events, it extracts trigger patterns, role definitions, and temporal constraints, then merges them with existing schemas, eliminating redundancy. The newly induced schemas are stored back into the MKB, closing the loop.

Finally, a transactional integration step atomically applies the knowledge increment ΔGₖ (new entities, new facts, and deprecated facts) to the current graph Gₖ₋₁, producing the updated graph Gₖ without requiring a full reconstruction.

The authors evaluate DIAL‑KG on both static benchmark datasets and a purpose‑built streaming corpus. Compared with strong schema‑free LLM baselines, DIAL‑KG improves F1 scores by up to 4.7 % and achieves over 98 % precision on evidence‑backed soft deprecations in the streaming setting. Moreover, the induced schemas are more compact, containing up to 15 % fewer relation types and reducing redundancy by 1.6–2.8 points, which translates into lower memory consumption and faster downstream reasoning.

Key contributions of the work include:

• A closed‑loop, batch‑granular incremental KG construction pipeline that integrates extraction, validation, and schema evolution.
• The dual‑track extraction mechanism that balances parsimony (using triples where sufficient) with temporal fidelity (using events only when necessary).
• The Meta‑Knowledge Base that functions as a self‑evolving constraint repository, enabling schema‑free yet structured graph growth.

Limitations noted by the authors involve the dependence on LLM context length for schema retrieval and the computational cost of repeated clustering and alignment at scale. Future directions suggest more efficient schema indexing, multi‑modal evidence incorporation, and exploration of domain‑specific fine‑tuning to further reduce hallucinations.

Overall, DIAL‑KG presents a compelling solution for dynamic, real‑time knowledge‑graph construction, offering a practical pathway to maintain high‑quality, up‑to‑date graphs without the overhead of periodic full‑graph rebuilds.