WiKC Project: Enhancing Knowledge Integrity

• WiKC Project is a comprehensive initiative that restructures collaborative knowledge systems, focusing on integrity, reliability, and dynamic maintenance.
• It employs advanced LLM-driven methods to refine taxonomies, eliminate redundancy, and ensure semantic accuracy in large-scale data graphs.
• The project develops risk observatories and iterative knowledge compilation workflows that enhance real-time QA performance and multilingual scalability.

The WiKC Project encompasses a suite of research and engineering initiatives aimed at restructuring, evaluating, and monitoring large-scale collaborative knowledge systems. Spanning domains such as taxonomy refinement using LLMs for Wikidata, risk observatories for knowledge integrity in Wikipedia, and knowledge compilation workflows for high-performance LLM-based retrieval, WiKC centers on foundational challenges of completeness, reliability, and dynamic maintenance in knowledge-centric platforms. The project’s contributions include automated taxonomy cleaning, observability for threat detection, iterative LLM wiki compilation, and scalable multilingual content architectures.

1. Foundations and Key Objectives

WiKC originated as the Wikipedia Knowledge Integrity Risk Observatory—conceived as an open, continuously updatable monitoring system to surface and quantify threats to the reliability, neutrality, and comprehensiveness of Wikipedia projects. The core, intertwined goals are:

• Developing comprehensive, hierarchical taxonomies of threats to knowledge integrity.
• Operationalizing these taxonomies via transparent, language-agnostic indicators that are cross-wiki comparable, language-independent, and periodically updatable.
• Laying the groundwork for an accessible dashboard supporting exploration by language, topic, and page granularity (Aragón et al., 2021).

WiKC’s research trajectory has since expanded, integrating advanced computational techniques for automated graph taxonomy cleaning in Wikidata via LLM-driven looped refinement (Peng et al., 2024), and designing knowledge-compilation frameworks (e.g., WiCER) for delivering sub-second, fully-grounded LLM-based question answering without retrieval failures (Huerta, 8 May 2026).

2. Taxonomy Refinement and Semantic Graph Cleaning

Automated refinement of Wikidata’s taxonomy addresses instance–class ambiguity, cycles, redundancy, inconsistent paths, and sparsity. The WiKC pipeline is formally defined on the subclass graph $G=(C,E)$:

• Acyclicity: Remove directed cycles.
• Minimal Redundancy: Eliminate edges where indirect transitive paths already exist, i.e., ensure $E_{tr} = \{ (u,v) \in E \mid \nexists$ path $u \to \cdots \to w \to \cdots \to v$ of length $\geq 2 \}$.
• Semantic Soundness: Preserve only those subclass relationships verifiable by human-readable labels/descriptions.
• Conciseness: Focus on well-described, densely populated classes.

Refinement is achieved through a loop combining graph mining, rule-based initial pruning, and LLM-based edge evaluation. Each edge is prompted in zero-shot mode through open-source LLMs (e.g., Mixtral-8×7B-Instruct-v0.1) to classify relations (subclassOf, superclassOf, equivalent, irrelevant, none). Based on this, the system executes cut, merge, reverse, or retain operations, then applies transitive reduction and final filtering for non-informative classes (Peng et al., 2024).

Intrinsic evaluation (complexity, conciseness, understandability) and extrinsic evaluation (entity typing accuracy at increasing depth) demonstrate order-of-magnitude gains: 4.1M $\to$ 17K classes, 4.8M $\to$ 20K edges, 35 $\to$ 0 cycles, and marked accuracy improvements (macro: 43% $\to$ 70%).

3. Knowledge Integrity Risk Assessment and Monitoring

WiKC’s observatory framework offers a systematic taxonomy of knowledge integrity risks, bifurcating them as internal (originating in the Wikimedia ecosystem) and external (originating outside Wikipedia). Major risk categories and representative indicators include:

Internal:

• Community Capacity: Ratio of editors with elevated rights ($R_1 = |E|/|A|$).
• Governance: Availability/effectiveness of dispute resolution and protections.
• Demographics: Editor socio-geographic diversity.
• Verifiability: Citation needed tags, unreliable sources.
• Quality: Stub/article ratios, class distributions.
• Controversiality: Edit wars, protection events, revert chains.

External:

• Media: Referrer-based traffic spikes ($Z = (V_t - \mu)/\sigma$).
• Geopolitics: Targeted edit campaigns and content skew.

Indicators are defined to be computable from standard Wikipedia data dumps, page-view logs, and API feeds. Shannon entropy is used for geo-diversity: $E_{tr} = \{ (u,v) \in E \mid \nexists$0, where higher $E_{tr} = \{ (u,v) \in E \mid \nexists$1 signals greater geographic spread (Aragón et al., 2021).

A case study utilizing edit and view entropy (2018–2021 logs) highlights that misalignment between editorial and readership origins (e.g., low S_views in Japanese Wikipedia despite varied edit sources) is correlated with increased vulnerability to misinformation, and entropy misalignment is an early-warning gauge of "coverage fragility."

4. LLM-driven Knowledge Compilation and QA Systems

WiKC advances LLM-based wiki memory artifacts to deliver context access at sub-second latency with zero retrieval failure, building on the "LLM Wiki" and Cache-Augmented Generation patterns. However, naive large-context compilation (“blind compilation”) causes severe information loss due to over-compression ("compilation gap"), with answer quality plunging and catastrophic (score=1) failures spiking from 17% to 53–60% as the LLM compiler discards critical facts required for accurate downstream responses (Huerta, 8 May 2026).

The WiKC centerpiece in this domain is WiCER (Wiki-memory Compile, Evaluate, Refine), an iterative CEGAR-inspired loop:

1. Compile raw documents into a condensed wiki artifact.
2. Evaluate against diagnostic QA probes (one per document); LLM judges return scores 1–5.
3. Extract minimal "pinned facts" from source articles for every catastrophic failure (score=1).
4. Recompile with explicit preservation of pinned facts; repeat until quality plateaus or failures vanish.

Empirically, WiCER recovers 80% of the quality lost to blind compilation and halves catastrophic failure rates (score=1 falls from 55.1% to 24.8%). Diagnostic pinning (as opposed to random constraint enforcement) is essential: ablative controls show less than one-sixth the improvement.

5. Systemic Multilinguality and Content Abstraction

An earlier foundational WiKC system addresses the multilingualism and coverage disparity in Wikipedia by splitting content and rendering logic:

• Abstract Wikipedia: A language-independent namespace for "constructors" and their arguments, expressing factual and structural content once globally.
• Wikilambda: A wiki of functions (stored with types, multiple implementations, and unit tests) that renders Abstract Wikipedia content into any target language at runtime, using information from Wikidata's lexemes for inflected forms, genders, etc. (Vrandečić, 2020).

This architecture allows:

• Centralized factual updates, reducing duplication.
• Synchronized, up-to-date content across all language editions.
• Community regulation of functions and type schemas.

The practical outcome is instant access to comprehensive, current material in low-resource language editions, and reduced risk of ideological lock-in via larger, more diverse co-editing.

6. Limitations and Forward Directions

Identified limitations across the WiKC initiatives:

• LLM Artifacts: Occasional hallucination of facts/relations, non-deterministic compilation behavior, and reliance on model quality and evaluation via LLM-as-judge (Peng et al., 2024, Huerta, 8 May 2026).
• Coverage: Taxonomy cleaning omits sparsely populated or highly domain-specific classes; knowledge compilation must balance brevity against completeness under context window constraints.
• Risk Metrics: Current indicators in observatory work are coarse-grained; finer per-page/category metrics, improved geo-attribution, and robust anomaly detection remain outstanding tasks (Aragón et al., 2021).
• Community Integration: Challenges include integration of semi-automated human-in-the-loop validation, harmonization with existing community workflows, and formalizing coverage/utility bounds for new downstream tasks.

Future proposals span hybrid Wiki-RAG inference employing retrieval fallback, adaptive per-document compression, and expanded evaluation for taxonomy coverage and real-world task performance (Peng et al., 2024, Huerta, 8 May 2026).

7. Implications and Applications

• Backbone Taxonomies: WiKC’s LLM-cleaned, redundancy-free taxonomies serve as high-precision backbones for tasks such as entity linking, alignment, and summarization.
• Community Health Monitoring: The observatory’s metrics provide actionable diagnostics for content quality, diversity, governance, and threat detection, increasingly important as Wikipedia remains a primary knowledge repository (Aragón et al., 2021).
• Efficient LLM QA: Knowledge compilation workflows ensure that models can operate at conversational latencies while maintaining answer faithfulness, overcoming the speed–completeness tradeoffs in classic RAG.
• Scalable Multilinguality: Abstraction of content and function unlocks near-universal access to knowledge, enabling rapid language expansion, consistency, and equity in global knowledge distribution (Vrandečić, 2020).

A plausible implication is that as these pipelines mature, combining fully abstracted multilingual architectures, robust risk observatories, refined taxonomy backbones, and iterative LLM-based knowledge artifacts could enable dynamic, reliable, and high-coverage knowledge systems unprecedented in both breadth and accuracy.