Personal Construct Theory and the Transfer of Human Expertise

John H. BooseAAAI 1984

TL;DR

This paper introduces the Expertise Transfer System (ETS), a novel approach to knowledge acquisition for expert systems that leverages George Kelly's Personal Construct Theory. ETS automates the interview process, eliciting and structuring an expert's problem-solving knowledge, significantly reducing the time and effort required for knowledge engineering.

What Problem Does This Paper Address?

Building expert systems is notoriously challenging due to the "knowledge acquisition bottleneck." Human experts often struggle to articulate their implicit, heuristic-based knowledge in a format that can be used by AI systems. Traditional knowledge engineering involves lengthy, iterative interviews between experts and knowledge engineers, which can be time-consuming (6-24 months for a prototype) and prone to misinterpretations or incomplete elicitation. This paper aims to streamline this process by automating the initial knowledge extraction and analysis.

Methods

The core of ETS is its application of George Kelly's Personal Construct Theory. This theory posits that individuals understand the world by forming "constructs" – bipolar dimensions of meaning (e.g., "hot" vs. "cold"). ETS uses a structured interview process inspired by Kelly's Repertory Grid technique:

  • Element Elicitation: The system first asks the expert to list "elements" (e.g., database types, diagnostic categories) that represent potential conclusions or classifications. This can be done in a list mode or an incremental mode.
  • Construct Elicitation: The system then presents triads of elements and asks the expert to identify a distinguishing trait (a construct) and its opposite that applies to two elements but not the third. This process is repeated to build a set of constructs.
  • Rating Grid Construction: The expert then rates each element against each construct on a numerical scale (or using "N" for neither, "?" for both). This forms a rating grid.
  • Analysis and Refinement: ETS employs non-parametric factor analysis to cluster similar constructs and generates an "entailment graph" to show implication relationships between constructs. This graph helps identify potential conflicts or inconsistencies in the expert's thinking. The system facilitates a dialogue with the expert to resolve these conflicts, either by editing ratings, adding new elements, or refining constructs.
  • Rule Generation: Based on the structured knowledge, ETS automatically generates heuristic production rules (conclusion rules and intermediate rules) with associated certainty factors, suitable for expert system building tools like KS-300™ and OPS5.

Datasets

The paper does not describe specific datasets used for training or evaluation in the traditional machine learning sense. Instead, the "dataset" is the knowledge elicited directly from the human expert during the interview process. The example session focuses on building a knowledge base for a "Database Management System Advisor."

Key Insights and Ideas

  • Personal Construct Theory as a Knowledge Engineering Tool: The paper demonstrates that psychological theories of human cognition can be effectively adapted for AI knowledge acquisition, offering a structured and less intimidating way for experts to articulate their knowledge.
  • Automated Interviewing Reduces Bottleneck: By automating the initial elicitation and analysis, ETS significantly cuts down the time and human effort required, making expert system development more accessible.
  • Leveraging Inconsistencies: Instead of forcing experts to immediately resolve inconsistencies, ETS uses them as points of exploration to uncover deeper aspects of the expert's problem-solving model and potentially refine their own understanding.
  • Rapid Prototyping: The system's ability to quickly generate a functional prototype knowledge base allows for early testing and analysis, accelerating the iterative development cycle.
  • Expert Enthusiasm: The paper notes that experts are generally enthusiastic about using ETS, finding it a valuable tool for organizing their thoughts and understanding their own decision-making processes.

Key Results

  • ETS can automatically produce knowledge bases for expert system building tools like KS-300™ and OPS5.
  • Knowledge engineers associated with ETS projects report saving 2-5 calendar months of knowledge acquisition time per project.
  • The system facilitates rapid prototyping, with an example of a Database Management Advisor knowledge base being generated in approximately two hours.
  • The process helps experts refine their thinking and identify potential conflicts in their problem-solving methods.

Limitations and Future Work

  • Depth of Knowledge: The paper acknowledges that ETS is best suited for analytic problems solvable by production systems and may not readily handle synthesis-class problems. Eliciting deep causal, procedural, or strategic knowledge remains a challenge, though alternate interviewing techniques are being explored.
  • Completeness of Constructs: It is difficult to verify if a sufficient set of constructs has been elicited. While inappropriate constructs can be weeded out, errors of omission are harder to detect.
  • Expert's Understanding: The methodology assumes the expert understands the problem's conclusion set. If not, the relevant set needs to be built through ETS's expansion methods or manual interviewing.
  • Future Work: Enhancements considered include expanding interview methods, incorporating more analytic tools for construct validity, developing feedback paths between ETS and target expert system tools, and exploring multi-dimensional scaling techniques. A knowledge engineering guide is also planned.

Summary updated

BibTeX
@inproceedings{boose1984personal,
  author = {John H. Boose},
  title = {{Personal Construct Theory and the Transfer of Human Expertise}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1984},
}
Figure