TL;DR
This paper introduces EXPECT, a knowledge acquisition (KA) system that uses explicit representations of problem-solving strategies to offer more flexible and adaptive KA. Unlike traditional "role-limiting" approaches that rigidly define KA tasks, EXPECT dynamically derives what knowledge is needed based on the explicit strategy, allowing for strategy modification and richer knowledge acquisition.
What Problem Does This Paper Address?
Building knowledge-based systems (KBS) requires acquiring knowledge from domain experts. A significant challenge in this process is Knowledge Acquisition (KA). Traditional KA tools often rely on "role-limiting" approaches, where a predefined problem-solving strategy (like "propose-and-revise") dictates specific roles that domain knowledge must fill. While this provides strong guidance, it lacks flexibility. The problem-solving structure of an application might not fit neatly into a predefined strategy, or a single strategy might not cover all aspects of a complex domain. Furthermore, these tools are often rigid, making it difficult to adapt the strategy itself or acquire knowledge beyond the predefined roles. This paper addresses this inflexibility by proposing a system, EXPECT, that represents problem-solving strategies explicitly, enabling dynamic adaptation and more comprehensive KA.
Methods
EXPECT's core innovation lies in its explicit representation of problem-solving strategies. This allows the KA tool to reason about the strategy and dynamically identify knowledge gaps.
- Explicit Problem-Solving Strategy Representation: Unlike role-limiting approaches where the strategy is implicit in the KA tool's structure, EXPECT represents the strategy in a procedural-style language integrated with a description logic knowledge representation system (LOOM). This representation includes methods, goals, and subgoals.
- Dynamic Knowledge Gap Identification: By analyzing the explicit strategy representation, EXPECT can detect missing information or inconsistencies. This is achieved through various analyzers (parser, method analyzer, instance analyzer) that identify errors and knowledge gaps.
- Flexible KA Guidance: When an error or gap is detected, EXPECT provides the user with suggested corrections. The system's guidance adapts as the knowledge base changes, offering a more interactive and context-aware KA experience.
- Unified KA Mechanism: A single KA tool in EXPECT can support any explicitly represented problem-solving strategy, eliminating the need to build a new tool for each strategy.
Datasets
Not discussed in the paper. The paper uses a U-Haul® rental domain example for illustration, which was adapted from existing work.
Key Insights and Ideas
- Explicit Strategy Representation is Key to Flexibility: The ability to represent and reason about problem-solving strategies explicitly is the fundamental enabler for flexible KA. This allows the system to understand why certain knowledge is needed, not just what knowledge is expected.
- KA Tool Adapts to Strategy, Not Vice-Versa: Instead of forcing the user to fit their knowledge into a predefined KA tool structure, EXPECT's KA tool adapts its requests and guidance based on the explicitly defined problem-solving strategy.
- Interdependencies Between Factual and Problem-Solving Knowledge: The paper highlights how changes in one type of knowledge (e.g., adding a new constraint) can necessitate changes or additions in the other (e.g., requiring a new method to check that constraint). EXPECT's explicit representation helps manage these interdependencies.
- Beyond Filling Roles: EXPECT goes beyond simply helping users fill predefined knowledge roles. It supports acquiring additional knowledge needed for problem-solving, including adapting the problem-solving strategy itself.
Key Results
The paper demonstrates EXPECT's capabilities through examples, particularly contrasting it with the role-limiting tool SALT for the "propose-and-revise" strategy.
- Acquiring Domain-Specific Factual Knowledge: EXPECT can request factual knowledge (e.g., capacity of a new truck model) that is directly relevant to the problem-solving strategy, even if it's not part of a predefined schema.
- Acquiring New Constraints and Fixes: When a new constraint is added, EXPECT can detect the need for a corresponding method to check it, illustrating its ability to manage the relationship between factual and procedural knowledge.
- Modifying Problem-Solving Strategies: The paper shows how EXPECT can support users in modifying the "propose-and-revise" strategy itself (e.g., adding constraint prioritization based on variable preference) and then automatically detecting the resulting knowledge gaps.
- Unified KA for Diverse Needs: EXPECT's single KA tool can handle various KA tasks, from acquiring input parameter values to modifying the core problem-solving strategy, demonstrating its versatility.
Limitations and Future Work
- User Choice and Context: While EXPECT suggests corrections, the paper notes that the user's choice of which suggestion to follow depends on the context. The system could potentially offer more sophisticated reasoning to guide this choice.
- User Roles: The paper mentions that supporting different user roles (e.g., domain expert vs. knowledge engineer) would require mechanisms to associate specific changes with specific user types.
- Bootstrapping with Existing Libraries: The authors suggest that EXPECT could benefit from representing existing libraries of methods to bootstrap the creation of new knowledge-based systems.
- Accessibility for End-Users: While the authors believe EXPECT's guidance is accessible to end-users, further work might be needed to ensure this across all levels of complexity.
Summary updated