Abstract: This paper presents a framework for "causal reconstruction," the process of building functional internal models of physical systems from sparse natural language descriptions. The developed system, PATHFINDER, maps simplified English into a "transition space" representation that tracks attribute changes over time. By utilizing exploratory transformations and Gricean conversational maxims, the system can bridge abstraction gaps and answer complex "how" and "why" questions about physical processes.
Introduction: The Challenge of Causal Reconstruction
Humans frequently learn the mechanics of complex systems through textual descriptions in manuals or encyclopedias. These descriptions are rarely exhaustive, relying on the reader's ability to infer missing links and map abstract concepts to physical realities. For AI, this presents a significant challenge, as traditional qualitative modeling often requires complete, pre-defined scientific models. This research addresses the gap by enabling a program to reconstruct the causal model intended by a writer from a sparse linguistic input.
The Transition Space Representation
The core of the PATHFINDER system is the Transition Space representation. Unlike static state modeling, this approach focuses on changes, depicting events as "path fragments" that track attribute transitions over ordered time points.
As illustrated in Figure 1, the system utilizes a specific set of primitives to categorize these changes. Boolean attributes are tracked via APPEAR and DISAPPEAR, while qualitative and quantitative attributes are monitored through CHANGE, INCREASE, and DECREASE (along with their respective negations). PATHFINDER, implemented in Common Lisp, uses a context-free grammar parser to translate simplified English into these event traces.
Bridging Gaps with Exploratory Transformations
A key innovation of this work is the use of "exploratory transformations" to align events that do not perfectly match in the initial description. These transformations include:
- Information-preserving transformations: Swapping equivalent attributes (e.g., "sliding" vs. "spinning").
- Non-information-preserving transformations: Generalization and "object composition." For instance, recognizing that an interaction with a specific component (like silver halides) constitutes an interaction with the larger object (the film).
The system also incorporates Paul Grice’s "Maxims of Conversation" as a formal framework for model evaluation. For example, the Maxim of Quantity ensures that the reconstructed model does not introduce extraneous objects not motivated by the text.
Case Study: Modeling State Changes
The system's ability to handle abstraction and object composition is vital for resolving the discontinuities common in human writing.
In the condensation example shown in Figure 4, the system tracks the transition of steam from vapor to liquid. Between time points t22 and t23, the is-a attribute for vapor "disappears" (D) while the is-a attribute for liquid "appears" (A). Simultaneously, the system maintains the state of contact and the absence of distance between the steam and the metal plate. This granular tracking allows the AI to confirm consistency even when a text omits intermediate physical steps.
Evaluation and Results
PATHFINDER was evaluated on a primary case study involving a camera's operation and over 60 additional causal descriptions covering domains such as combustion, chemical reactions, and radio signals. The system demonstrated the ability to answer four distinct types of questions:
- Time-varying status: Tracking changes like the decreasing distance between light and film.
- Inter-event relationships: Linking separate events, such as light convergence and image formation.
- Plausible causal connections: Identifying the chain of events leading from a reflected image to a physical contact on film.
- Consistency checking: Determining if new statements (e.g., "Does the light contact the emulsion?") are consistent with the reconstructed model.
Conclusion
This research distinguishes "causal reconstruction" from traditional modeling. While modeling builds from first principles, reconstruction seeks to recover a specific model held by a writer using text as a blueprint. By grounding physical primitives in language, PATHFINDER bridges the gap between linguistic input and qualitative physical reasoning. Future work will focus on handling more complex phenomena, such as feedback cycles and repetitive events.
Summary updated