Modern intelligent systems face a persistent tension: they are most effective when grounded in structured, expert-verified knowledge, yet this very grounding makes them brittle and difficult to scale. Across decades of research, a recurring pattern emerges where the integration of domain-specific ontologies and knowledge graphs (KGs) provides a massive performance boost but creates a "manual bottleneck" that prevents these systems from moving easily between disciplines.
The Power of Knowledge Infusion
To move beyond simple pattern matching, researchers have long sought to "infuse" AI with structured logic. By tethering a model to a domain-specific knowledge base, developers can transform a passive system into a proactive one.
For example, Yang Deng and colleagues developed KEMI, a dialogue framework that uses mental health knowledge to transition from empathetic listening to proactive problem-solving Deng et al.. Similarly, the D3LM system by Yang Wu and others uses legal knowledge graphs to allow Large Language Models (LLMs) to act like lawyers, identifying missing facts in a user's story to provide more accurate legal "Court Views" Wu et al..
In these cases, the structured knowledge acts as a "guardrail" and a "reasoning engine," ensuring the AI stays grounded in the specific rules and facts of a field—whether that is the nuances of classical Chinese poetry Li et al. or the complex data structures of computer science Kersey et al..
The Manual Reconfiguration Trap
The intuition behind why these systems work is clear: structured knowledge provides the "common sense" or "expert logic" that raw data lacks. However, the evidence across the literature reveals a shared failure mode: domain dependency.
As Massimo Poesio and Claudio Rullent noted as early as 1987, architectures that integrate syntactic and semantic rules into pre-compiled sources suffer when transitioning to new subject areas, requiring a complete re-compilation of the knowledge base Poesio and Rullent. This sentiment is echoed in modern work on image generation Aivalis and scientific claim generation Wright and Wadden, where the effectiveness of the system is strictly constrained by the availability of high-quality, expert-designed ontologies.
This creates a "scalability bottleneck." If a system designed for biomedical evidence discovery Wang et al. needs to be moved to the legal or engineering domain, it cannot simply be retrained on new text. Instead, it requires the labor-intensive construction of new, specialized knowledge structures—a process that Jos de Bruin and Remko Scha identified as a significant manual hurdle for translating linguistic structures into formal queries de Bruin and Scha.
Toward Automated Knowledge Elicitation
To break this bottleneck, research has shifted toward systems that can "learn how to learn" a new domain. The goal is to move from static knowledge bases to active ones.
- Self-Gapping: Systems like KRITON use an active monitoring agent to detect gaps in their own understanding and trigger the most effective elicitation method to fill them Diederich.
- Distant Supervision: Tools like EvidenceMiner use meta-patterns and automated distant supervision to find evidence in literature without requiring manually labeled training data Wang et al..
- Unsupervised Adaptation: Researchers like Bo Chung have developed ensemble methods that use genetic algorithms to tune term extraction for new medical specialties without massive manual labeling Chung et al..
The Future of Cross-Domain Reasoning
The ultimate objective of this research direction is to create systems that possess the precision of a knowledge graph with the flexibility of a general-purpose model. Recent efforts like CS-KG 2.0 Dessì et al. and the UKnow protocol Tan et al. attempt to build massive, multi-million-node graphs to provide a "knowledge-view" backbone for AI.
However, as Menasha Thilakaratne and Katrina Falkner argue, true innovation lies in Literature-Based Discovery (LBD)—the ability to identify hidden associations across siloed research fields Thilakaratne and Falkner. By moving toward domain-agnostic frameworks that can automatically extract "functional facets" (like tasks and methods) rather than just keywords Horvitz, the field is slowly dismantling the manual barriers that have historically kept expert AI locked within narrow silos.