TL;DR
This paper tackles the challenge of generating natural language explanations from massive, complex knowledge bases. The authors introduce KNIGHT, a system that successfully generates detailed explanations, and a novel "Two-Panel Evaluation Methodology" to rigorously assess its performance against human experts. Their findings demonstrate that AI can produce explanations comparable in quality to those of domain experts, paving the way for more sophisticated AI explanation systems.
What Problem Does This Paper Address?
Generating natural language explanations for complex phenomena is a crucial capability for AI systems, enabling them to communicate knowledge effectively. While significant progress has been made in developing computational mechanisms for explanation generation, a major hurdle has been the lack of robust empirical evaluation. This is largely due to the absence of large-scale, semantically rich knowledge bases that can serve as the foundation for such systems and provide sufficient data for meaningful studies. Existing research often relies on "proof-of-concept" systems tested on limited examples, making it difficult to assess scalability and real-world applicability. This paper aims to bridge this gap by empirically evaluating an explanation system built upon a substantial knowledge base.
Methods
The core of this research lies in the development and evaluation of the KNIGHT system. KNIGHT is designed to:
- Extract information: It accesses a large, formal representation of domain knowledge.
- Organize information: It structures the extracted knowledge into multi-sentential discourse plans.
- Realize information: It translates these plans into coherent, multi-paragraph text.
Key methodological innovations include:
- Large-Scale Knowledge Base: The authors utilized the "Biology Knowledge Base," a massive repository containing over 180,000 facts about botanical anatomy, physiology, and development. This provided a rich source of information for generating complex explanations.
- KNIGHT System Architecture: KNIGHT employs a modular architecture that separates knowledge-base access from explanation planning. This separation is crucial for robustness and scalability. It uses "KB Accessors" to retrieve "views" (coherent subgraphs) of concepts from the knowledge base.
- Explanation Design Packages (EDPs): These are frame-based structures that encode discourse knowledge, providing abstractions for specifying the content and organization of explanations. EDPs combine frame representations with procedural constructs, mirroring the structure of expository texts and facilitating reuse and modification.
- Two-Panel Evaluation Methodology: This novel approach was designed to rigorously assess explanation quality. It involves:
- Generating explanations using the computer system (KNIGHT).
- Forming two panels of domain experts: a "Writing Panel" to generate human explanations and a "Judging Panel" to evaluate all explanations.
- Ensuring "Computer Blindness" – judges were unaware of the source of the explanations (human or machine).
- Evaluating explanations based on criteria like overall quality, coherence, content, organization, writing style, and correctness.
Datasets
The primary dataset used in this study is the Biology Knowledge Base. This is a newly constructed, large-scale knowledge base containing over 180,000 facts about botanical anatomy, physiology, and development. It was not specifically tailored for explanation generation, which adds to the robustness of the evaluation.
For evaluation, 388 concepts were initially selected from the Biology Knowledge Base. After applying a "length filter" (requiring at least 3 sentences), 87 explanations were retained (48 object descriptions, 39 process descriptions). Finally, 30 object and 30 process explanations were randomly chosen for the evaluation, resulting in a total of 60 explanations generated by KNIGHT.
Key Insights and Ideas
- Scalability is Achievable: The research demonstrates that it's possible to scale explanation generation systems to large, complex knowledge bases, moving beyond small-scale demonstrations.
- Separation of Concerns is Key: Decoupling knowledge-base access from explanation planning leads to more robust systems that can handle diverse and potentially irregular knowledge representations.
- Rigorous Evaluation Matters: The "Two-Panel Evaluation Methodology" provides a strong framework for objectively comparing AI-generated explanations against human experts, addressing the critical need for empirical validation in NLG.
- Human-Level Performance is Attainable: The results suggest that AI can generate explanations that are nearly indistinguishable in quality from those produced by human domain experts, particularly in areas like content and correctness.
- Discourse Knowledge Engineering is Crucial: The development of EDPs highlights the importance of structured, reusable representations for discourse knowledge to effectively guide explanation planning.
Key Results
The evaluation yielded highly encouraging results:
- Overall Quality: KNIGHT scored within "half a grade" of the human experts on overall quality and coherence.
- Specific Dimensions:
- KNIGHT showed no statistically significant difference from human experts in content, organization, and correctness.
- KNIGHT performed best on correctness and content, and least well on organization.
- Individual Writer Comparison: KNIGHT's performance was comparable to or exceeded that of three out of four individual human writers. Specifically, KNIGHT was very close to "Writer 3" and actually exceeded the performance of "Writer 4."
Limitations and Future Work
- Limitations:
- The paper acknowledges that while differences were not statistically significant in some dimensions, a larger sample size might reveal more subtle differences.
- The "Computer Blindness" was maintained, but the authors note that the judges were not aware of the purpose of the experiment, implying potential for bias if they suspected the AI's involvement.
- The KB accessing system was minimally addressed in terms of its query interpreter capabilities.
- Future Work:
- The authors suggest that the developed methodology and system architecture can be extended to other domains.
- Further research could explore more sophisticated user modeling to tailor explanations even further.
- The work lays the groundwork for developing explanation systems that can engage in more complex interactive dialogues.
Summary updated