Traditional knowledge representation is built on the "centralized database" mindset: the goal is to resolve contradictions and arrive at a single, globally consistent source of truth. However, as Mind the Web argues, this approach fails at web-scale, where information is decentralized, autonomous, and inherently inconsistent. In domains like history, culture, and personal taste, forcing a single consensus doesn't just simplify the data—it destroys the very nuance that makes the data meaningful.
Researchers are moving toward polyvocal knowledge representation, a paradigm where contradictions are preserved as features rather than filtered out as noise.
The Distinction Between Fact and Taste
The foundation of this shift lies in distinguishing between different types of knowledge. Walter Quattrociocchi and Dino Karabeg in Taste and Trust argue that while we can seek a single undisputed truth for "taste-independent" facts (like a mechanic’s technical competence), we must accommodate "taste-dependent" opinions (like movie preferences).
When systems treat subjective knowledge as if it were objective, they suffer from "thinness." As Cody Kommers and colleagues suggest in Meaning Is Not A Metric, numerical metrics strip away the cultural context that gives human experiences their "thickness." To capture this, we need frameworks that allow multiple, potentially conflicting interpretations of the same event to coexist.
Reifying Interpretation as a First-Class Entity
The technical mechanism for achieving this is reification: treating the act of interpretation itself as a data object.
In the VAST Research Project, Alfio Ferrara, Silvana Castano, and their team model the relationship between a source text, a specific value (like "freedom"), and the historical context or scholar providing the analysis. By making the "interpretation" a first-class entity in the database, the system can store two contradictory scholarly views of the same text without triggering a logical error.
This approach is echoed in Event-Model-F, where Ansgar Scherp and colleagues separate a real-world occurrence from its various interpretations. This allows an emergency response system, for example, to store different stakeholders' theories about the cause of an incident simultaneously.
Cultural Nuance and the Failure of Universality
The need for polyvocality is most visible in cultural and "commonsense" knowledge. Quyet V. Do and others point out in What Really is Commonsense Knowledge? that commonsense is not a universal constant; it is a subjective set of beliefs that varies across geography and time.
Consider culinary traditions. Li Zhou and colleagues found that Large Language Models (LLMs) exhibit a strong Western bias, often failing to recognize regional variations in dishes (Does Mapo Tofu Contain Coffee?). Similarly, Weiqing Min notes in The Development and Applications of Food Knowledge Graphs that food data often involves conflicting health beliefs.
If a knowledge graph forces a single "correct" recipe or health claim, it loses the cultural diversity documented by Paul Laufer and Markus Strohmaier, who showed that Wikipedia's portrayal of food is filtered through specific cultural lenses (Mining cross-cultural relations from Wikipedia).
Preserving Disagreement as a Feature
How do we build systems that don't collapse under the weight of these contradictions? The emerging strategy is to treat disagreement as provenance metadata.
- Preserve Divergence: In crowdsourced systems, Jinta Weng suggests that if user consensus falls below a threshold, the system should preserve all conflicting relationships rather than picking a "winner" (Construction and Application of Teaching System Based on Crowdsourcing Knowledge Graph).
- Bridge Subjective Gaps: In mental health, Akshata Kishore Moharir uses heterogeneous graphs to bridge the gap between how patients express distress in native languages and Western clinical categories, preserving alternative interpretations as valid paths (Cross-Lingual Mental Health Ontologies for Indian Languages).
- Benchmark for Nuance: New benchmarks like PalmX 2025 by Fakhraddin Alwajih and colleagues are being designed to test LLMs specifically on their ability to handle the fragmented, non-canonical realities of diverse cultures.
By adopting polyvocality—as advocated in the SEMMES 2024 workshop by Pasquale Lisena and Simon Gottschalk 0001—data models move beyond static, objective facts. They become capable of representing the multi-vocal narratives that define human history, culture, and identity.