Abstract: In the 1980s, Artificial Intelligence research was polarized between "neat" logical approaches and "scruffy" heuristic methods. Eugene Charniak’s Wimp (Wholy Integrated Marker Passer) attempted a synthesis, using marker passing as a heuristic to find potential logical proofs for story understanding. This paper outlines the Wimp architecture, its use of abductive reasoning, and its evaluation via a predictive "figure of merit."
Introduction
The mid-80s AI landscape was divided. "Scruffy" researchers favored spreading-activation networks (marker passing) for their speed and intuitive handling of associations, but these systems lacked formal rigor. "Neat" researchers insisted on first-order logic, which provided rigor but suffered from combinatorial explosion. Charniak proposed Wimp to unify these paradigms, using marker passing to identify "skeletal proofs" that are subsequently validated and fleshed out by a logical path checker.
System Architecture
Wimp integrates several modules to process natural language. Morphology and syntax provide initial structural analysis, while the Marker Passer explores the Knowledge Representation (KR) to find semantic connections. The Path Checker acts as the bridge, evaluating the paths found by the marker passer against the logical constraints of the KR.
The Wimp Framework
Marker Passing and Knowledge Representation
When a word is encountered, "marks" spread from its node in the knowledge graph. These marks carry "zorch"—a measure of activation strength that decays with distance. When marks from different concepts collide, a potential path is identified. This process allows the system to quickly find relevant connections in a large network of facts.
Path Checking and Abduction
Wimp treats identified paths as the "backbone" of a logical proof. It employs abductive reasoning—making minimal assumptions to complete a logical explanation. For instance, if "Jack" and "supermarket" are linked, the system might assume "Jack is shopping" to explain the connection. This turns a simple association into a structured explanation.
Evaluation via Figure of Merit
To choose between competing explanations, Wimp uses a "figure of merit" based on predictive power:
Figure of Merit = (Number of True Predictions) - (Number of Assumptions)
A superior explanation is one that accounts for more observed text (high predictive power) while requiring fewer unverified leaps of faith (low number of assumptions).
Key Insights and Results
Wimp demonstrated that word-sense disambiguation and reference resolution could be unified under a single abductive mechanism. By using marker passing to prune the search space, Wimp avoided the computational pitfalls of traditional theorem provers while maintaining logical consistency. Qualitative tests on short story snippets showed the system could successfully infer character motivations and resolve ambiguous references, such as determining that a "rope" in a specific context is intended for a noose.
Limitations and Future Directions
While successful on a small knowledge base of approximately 225 facts, the scalability of marker passing remains a concern; larger networks might generate excessive "garbage" paths. Furthermore, Charniak noted that the binary nature of logical assumptions might eventually need to be replaced by a more nuanced probabilistic or Bayesian framework to better represent the varying strengths of different inferences.
Summary updated