Introduction
Extracting qualitative physical and causal information from natural language is a significant challenge for artificial intelligence. While humans easily understand statements like "if the temperature rises, the pressure increases," translating such textbook explanations into formal representations often fails due to linguistic ambiguity and complex discourse structures. Traditional syntactic pattern matching is insufficient for capturing information spread across multiple sentences. This paper introduces a framework that utilizes "narrative functions"—the intended purpose of a sentence within a broader discourse—combined with abductive reasoning to translate natural language into formal Qualitative Process (QP) theory representations.
Methodology: Narrative-Aware Extraction
The core of the proposed system is the Explanation Agent Natural Language Understanding (EA NLU). It moves beyond simple syntax parsing by identifying the functional role of sentences within a narrative.
Narrative Functions and QP Frames
The system maps specific sentence types to narrative functions, which in turn guide the creation of QP frames. These frames represent quantities, influences, processes, and topological constraints. For instance, a sentence might serve the function of IntroductionOfQuantityEvent or IntroductionOfQPropEvent.
Abductive Reasoning
To resolve linguistic ambiguities, the system employs query-driven abduction. By setting top-down expectations based on narrative functions, the system biases its interpretations toward those that produce valid QP frames. This approach makes abductive reasoning computationally efficient by filtering out irrelevant interpretations that do not contribute to the qualitative model.
Type-Level Influences
A significant contribution of this work is the extraction of "type-level" causal relationships (e.g., "Adding a university increases science output"). These general rules allow the system to provide strategic advice to autonomous agents, moving from specific instances to broader conceptual understanding that can be applied across different scenarios.
Evaluation and Results
The system was evaluated across three distinct datasets: a foundational benchmark of 8 examples, a science text corpus, and strategic advice for the game Freeciv.
Performance on Science Texts
On a corpus of nine simplified paragraphs from a science book, the system achieved a precision of 0.74 and a recall of 0.45. While recall remains a challenge due to the sheer variety of human expression, the high precision indicates that the extracted qualitative frames are generally correct and reliable.
Enhancing Game Agents
The most compelling results came from the strategy game Freeciv. An AI agent provided with just six sentences of qualitative advice extracted by the system showed statistically significant improvements over a baseline agent.
The experimental agent demonstrated superior science output (p < 0.041) and population growth, proving that a small amount of high-level qualitative advice can be more effective for an AI than extensive computation or purely demonstration-based learning.
Conclusion and Future Work
By focusing on narrative functions, this research demonstrates a robust method for extracting actionable qualitative knowledge from text. While coverage remains a hurdle—particularly regarding coreference resolution and complex continuous processes—the current system successfully bridges the gap between natural language explanations and formal causal reasoning. Future work will explore integrating analogical reasoning to allow the system to learn from fewer examples by comparing new text to previously understood concepts.
Summary updated