From Philosophy to Interfaces: an Explanatory Method and a Tool Inspired by Achinstein’s Theory of Explanation

Document Title

From Philosophy to Interfaces: an Explanatory Method and a Tool Inspired by Achinstein’s Theory of Explanation

Document Information

• Subject Area: Explainable Artificial Intelligence (XAI) and Human-Computer Interaction
• Keywords: Methodology, Learning-related Technologies, Explainable Artificial Intelligence, Explanation Generation, User-Centered Design, Multilingual Knowledge Graphs
• Case Study: Comparative Experiment on an XAI-powered Credit Approval System

Research Background and Problem Statement

• Identified Problems or Challenges:
  1. Current XAI technologies often adopt a "one-size-fits-all" approach, providing fixed explanations that overlook the diversity in user backgrounds, needs, and goals.
  2. Existing explanation frameworks in complex systems fail to meet users' needs for personalized and easily understandable explanations.
  3. Most XAI technologies focus on defining explainability as making things "understandable," rather than providing user-centered, precise, and personalized explanation processes.
• Significance: Providing user-friendly, transparent, and comprehensible explanations in complex AI systems is critical. Not only do users need to trust AI decisions, but regulations like the EU GDPR also mandate the right to explanation.
• Research Motivation and Related Work: This paper proposes a user-centered explanation process model based on Achinstein's philosophical theory of explanation, transitioning from theory to a practical tool. Related work includes research on philosophical theories of explanation (e.g., Achinstein's studies) and existing static contrastive explanation tools (e.g., IBM's Contrastive Explanations Method, CEM).

Solution

• Proposed Methods and Approach:
  1. Introduced a model based on Achinstein's theory of explanation, defining explanations as answers to users' explicit and implicit questions.
  2. Hypothesized that by identifying a set of simplified prototypical questions, highly generalized "Explanatory Overviews" can be generated.
  3. Proposed an AI algorithm pipeline to construct knowledge graphs from unstructured natural language documents and generate interactive explanatory overviews based on these graphs.
• Innovations:
  1. Operationalized Achinstein's philosophical definition of explanation into software implementation, particularly by translating it into answers to a set of prototypical questions.
  2. Proposed a user-oriented explanation generation model that dynamically adjusts explanation paths based on user behavior (via the generation of an "Explanatory Space").
  3. Leveraged advanced statistical and natural language processing techniques to match questions and answers, ensuring relevance without relying on complex human intent modeling.
• Implementation Steps and Key Technologies:
  1. Knowledge Graph Generation: Extracted concepts and relationships from unstructured natural language to construct graphical representations, using tools like Spacy for syntactic dependency parsing.
  2. Information Classification and Summarization: Categorized knowledge graph information into prototypical questions (e.g., Why, What for, How) and used deep learning models for prioritization and summary generation.
  3. Visualization and Interactive Interface: Enhanced static explanation tools (e.g., CEM) with interactive features, allowing users to click on nodes to view multi-level overviews and dynamically explore related information.

Research Outcomes

• Specific Outcomes:
  1. Developed and tested a new method for generating user-centered explanations.
  2. Designed and implemented an interactive tool using this model in the IBM credit approval system case study.
  3. Demonstrated the feasibility of Achinstein's theory in user-centered explanation generation.
• Advantages over Existing Solutions:
  1. Compared to static CEM tools, the new method generates personalized explanations that are easier to explore and dynamically adjust.
  2. Improved the quality of information users received (reflected in the experiment's "effectiveness" scores).
• Experimental or Evaluation Results:
  1. In a user study with over 100 participants, the new system significantly outperformed the baseline system in effectiveness (U=931.0, p=0.036).
  2. No significant differences were observed in efficiency (time spent) and satisfaction, partly due to the increased exploration burden.
  3. Quantitative and subjective evaluation results showcased the improved effectiveness of the new explanation model, though users may face a learning curve.
• Limitations and Future Directions:
  1. Overly generic "prototypical questions" may increase the exploration burden for users; more specific user questions are recommended.
  2. Combining the current method with traditional natural language question-answering systems could further simplify user needs localization.
  3. Conducting experiments in more application domains to validate generalizability and exploring ways to reduce users' learning effort.

The structured summary above highlights the core concepts, technical innovations, and experimental validations presented in this paper.