Direct Elicitation of Expert Knowledge for Machine Reasoning

69 concepts65 papers2 proposals

To build an intelligent system capable of reasoning about the real world—whether diagnosing a medical condition, predicting mineral deposits, or evaluating enterprise risk—the system requires structured, formal knowledge. However, a fundamental friction exists at the heart of knowledge engineering: human expertise is inherently qualitative, nuanced, and frequently tacit, while computational systems demand rigid, unambiguous representations like Description Logics, ontologies, or mathematical constraints.

Historically, this mismatch created what is known as the knowledge acquisition bottleneck. As Brian Phillips et al. observed in An Eclectic Approach to Building Natural Languages Interfaces, relying on a human intermediary (such as a knowledge engineer or programmer) to translate an expert's mental model into code is highly inefficient and prone to misinterpretation. When domain experts cannot directly author or audit the system's underlying logic, the resulting knowledge bases become opaque and difficult to verify, as noted by Douglas R. Skuce in An English-Like Language For Qualitative Scientific Knowledge.

This translation gap spans multiple generations of AI technologies. In traditional expert systems, it manifested as the arduous task of writing manual production rules (Deborah L. McGuinness et al., Description Logic in Practice: A CLASSIC Application). In the era of the Semantic Web, it appeared as a modeling expertise gap where domain scientists struggled to grasp abstract ontological concepts like OWL or RDF (Soonho Kim et al., Software architecture for managing the Fisheries ontologies lifecycle; Ratan Bahadur Thapa et al., A Roadmap to Create a Knowledge Graph for the Circular Factory for the Perpetual Product). Even in modern machine learning, the challenge persists when attempting to encode clinical intuition into mathematical constraints for deep learning optimization (Jianwei Niu et al., A survey on incorporating domain knowledge into deep learning for medical image analysis) or integrating scientific principles into neuro-symbolic architectures (Weilin Chen et al., Exploring neuro-symbolic AI applications in geoscience: implications and future directions for mineral prediction).

To bypass the technical intermediary, researchers have developed frameworks that allow experts to directly interact with and shape the system's logic. These solutions rely on three core mechanisms: controlled natural languages, visual abstractions, and structured interactive elicitation.

Controlled Natural Languages and Feedback Loops

One prominent strategy for lowering the accessibility barrier is allowing experts to communicate with the system using a constrained subset of natural language. The intuition here is that if the system can restrict grammar to structures that map directly to formal logic, the expert can write naturally while the machine compiles the text into executable rules.

Early implementations, such as the KAL language proposed by Douglas R. Skuce, targeted qualitative scientific knowledge, allowing clinicians and biologists to express relationships without seeing formal predicates. To ensure accuracy, these systems introduced interactive feedback loops. For instance, the "WYSIWYM" (What You See Is What You Meant) approach introduced by Richard Power et al. in Multilingual authoring using feedback texts allowed users to edit knowledge bases through natural language feedback texts. By presenting the system's internal state back to the user in clear text, the expert could immediately spot and correct translation errors without needing specialized knowledge engineering skills.

Visual and Spatial Abstractions

Another powerful mechanism for bypassing technical complexity is shifting from textual logic to visual manipulation. Humans are naturally adept at spatial reasoning, and structures like mind maps or concept maps closely mirror how experts organize hierarchical information.

Early graphical tools like IMPULSE, developed by Eric Schoen et al. (IMPULSE: A Display Oriented Editor for STROBE), proved that multi-window visual editors could streamline knowledge base modification. This concept evolved into sophisticated web-based platforms like SHAKEN, where Jérôme Thoméré (A Web-Based Ontology Browsing and Editing System) demonstrated that domain experts could construct complex, logic-based knowledge bases entirely through graphical concept maps. Similarly, the AURA system (Ken Barker et al., AURA: Enabling Subject Matter Experts to Construct Declarative Knowledge Bases from Science Textbooks) combined visual mapping with automated reasoning engines to ingest textbook knowledge directly from educators.

The utility of visual abstraction extends to automated reasoning and planning. For example, in enterprise risk management, the Scenario Planning Advisor developed by Shirin Sohrabi et al. (An AI Planning Solution to Scenario Generation for Enterprise Risk Management) takes user-generated Mind Maps and automatically compiles them into formal AI planning languages. This allows business leaders to perform complex "what-if" reasoning over real-world data streams without ever interacting with rigid logical planning syntaxes.

Structuring the Elicitation Process

These approaches succeed because they replace open-ended programming with structured elicitation frameworks. Instead of asking an expert to write a program, these systems constrain the interaction to specific cognitive tasks that humans excel at, such as categorization, justification, and responding to targeted queries.

For example, Tom Bylander et al. demonstrated in CSRL: A Language for Expert Systems for Diagnosis that mapping raw, continuous data onto a symbolic scale helps mimic human diagnostic categorization, easing the knowledge acquisition bottleneck. From a formal learning perspective, Boris Konev et al. (A Model for Learning Description Logic Ontologies Based on Exact Learning) showed that complex Description Logic ontologies can be efficiently constructed by using an "Exact Learning" framework that poses simple, data-driven questions to a domain expert, mathematically minimizing the cognitive load required to build precise models.

The Modern Paradigm: Curation, NLP, and LLMs

As AI has shifted from symbolic rules to machine learning and Large Language Models (LLMs), the knowledge engineering bottleneck has not disappeared; it has merely changed form. Today, as Bill Howe argues in AI is Reducing to Data Curation (Abstract), we are entering a "curation as programming" paradigm where the primary challenge is enabling experts to inject their knowledge directly into data pipelines and model alignment workflows.

Modern frameworks apply structured elicitation to bridge the gap between domain experts and data scientists. The Ziva framework, introduced by Soya Park et al. (Facilitating Knowledge Sharing from Domain Experts to Data Scientists for Building NLP Models), moves beyond simple data labeling by prompting experts for "concept creation" and "justification elicitation." This converts qualitative expert logic into a reusable repository that data scientists can directly leverage for feature engineering and model debugging.

Similarly, in the context of generative AI, Ananjan Nandi et al. developed Roleplay-doh: Enabling Domain-Experts to Create LLM-simulated Patients via Eliciting and Adhering to Principles. Recognizing that mental health counselors possess deep intuitive knowledge of human behavior but lack prompt engineering skills, the authors designed a multi-stage pipeline that extracts qualitative feedback and converts it into structured natural language "principles." The system then enforces adherence to these principles, allowing non-technical experts to successfully "sculpt" highly authentic AI simulations.

Whether through the controlled natural languages of the 1970s, the visual concept maps of the 2000s, or the interactive prompt-sculpting frameworks of today, the core insight remains identical: overcoming the knowledge engineering barrier requires building interfaces that adapt to the expert's natural cognitive models, rather than forcing the expert to adapt to the machine's underlying logic.

Key Foundational Work

Meta-Level Knowledge: Overview and Applications (1977) — contributed 1 concept elements to this direction.
Interactive Transfer of Expertise: Acquisition of New Inference Rules (1977) — contributed 1 concept elements to this direction.
An English-Like Language For Qualitative Scientific Knowledge (1975) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_knowledge_acquisition_bottleneck_knowled_859a6db0,
  title = {knowledge acquisition bottleneck (knowledge engineering bottleneck)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/knowledge-acquisition-bottleneck-knowledge-engineering-bottleneck/rdr_859a6db07f144a64bf5943b17260cb29},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_knowledge_acquisition_bottleneck_knowled_859a6db0_v1,
  title = {knowledge acquisition bottleneck (knowledge engineering bottleneck)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/knowledge-acquisition-bottleneck-knowledge-engineering-bottleneck/rdr_859a6db07f144a64bf5943b17260cb29/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake