Scientific Explanation and Natural Language: A Unified Epistemological-Linguistic Perspective for Explainable AI

Marco Valentino, André FreitasarXiv 2022 · DOI · arXiv

Introduction

The rapid expansion of Explainable AI (XAI) has led to a proliferation of models capable of generating natural language explanations. However, these systems often operate without a rigorous theoretical foundation. While researchers typically evaluate explanations based on logical entailment or supporting evidence, they seldom address the fundamental epistemological question: what constitutes a scientific explanation? This paper bridges the gap between the Philosophy of Science and XAI by analyzing how formal explanatory theories manifest in natural language datasets.

Theoretical Framework: From Logic to Unification

To ground XAI in established epistemology, the authors review several major philosophical accounts of scientific explanation:

  1. Deductive-Nomological (DN) and Inductive-Statistical (IS): These models, pioneered by Hempel, view explanation as a logical argument where the explanandum (the event to be explained) is deduced from the explanans (initial conditions and universal or statistical laws).
  2. Unificationist Account: Proposed by Kitcher, this view suggests that explaining is the act of fitting specific events into broader, recurring patterns or "argument patterns."
  3. Causal-Mechanical (CM): This account focuses on citing the specific causal processes and interactions that lead to an event.
Hempel's formal models of explanation: Deductive-Nomological (left) and Inductive-Statistical (right).

Methodology and Corpus Analysis

The researchers tested these philosophical models against two primary natural language corpora:

  • Biology Why Corpus: A collection of 193 technical "why" questions and answers extracted from biology textbooks.
  • WorldTree: A semi-structured corpus of 2,206 science exam questions featuring manually curated explanation graphs designed for multi-hop reasoning.

The study employed quantitative metrics, such as fact reuse frequency, and qualitative annotation to determine which philosophical frameworks best describe human-generated explanations.

Key Insights: The Power of Unification and Abstraction

A central finding of the research is that explanations are not merely predictions. While one can predict a flagpole's height from its shadow, the shadow does not explain the height. This asymmetry is a hallmark of true explanation.

The authors argue that the primary function of a successful explanation is unification. By showing how a seemingly isolated event (e.g., a glass falling) fits into a universal regularity (e.g., gravity), the explanation provides deeper understanding.

The Unificationist account (Kitcher, 1989) illustrates how a theory subsumes a class of phenomena through argument patterns.

To achieve this unification, language must move from the concrete to the abstract. For instance, explaining why a "glass" falls requires categorizing it as an "object with mass." This makes taxonomic relations (IS-A) and definitions the "inference substrate" of explanation. In the WorldTree corpus, the most frequently reused explanatory facts—those with the highest "unification power"—were found to have the lowest lexical similarity to the questions they addressed, confirming their high level of abstraction.

Empirical Results

The analysis revealed that while causality is a dominant mode of explanation—75% of the Biology Why Corpus cited direct mechanisms—it is rarely sufficient on its own. Approximately 73% of explanations also included definitions or attributes, supporting the idea that human-centric explanation requires both causal links and categorical grounding. The study further identified 11 recurring knowledge categories, such as Taxonomic and Requirement, that serve as the building blocks of explanatory discourse.

Conclusion and Future Directions for XAI

This research suggests that XAI evaluation should shift from surface-level similarity metrics to assessments of a model's ability to perform abstraction and unification. By prioritizing high-level regularities, "unification-aware" models could significantly reduce the search space in complex reasoning tasks, leading to AI systems that provide more meaningful and scientifically grounded explanations.

Summary updated

BibTeX
@misc{valentino2022scientific,
  author = {Marco Valentino and André Freitas},
  title = {{Scientific Explanation and Natural Language: A Unified Epistemological-Linguistic Perspective for Explainable AI}},
  year = {2022},
  eprint = {2205.01809},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2205.01809},
}
Figure