Picnics, Kittens And Wigs: Using Scenarios For The Sentence Competition Task

C. Bullwinkle

Abstract

In the mid-1970s, natural language processing (NLP) struggled with "common sense" reasoning—the ability to infer situational context beyond literal syntax. This paper presents a system implemented in LISP that utilizes structured "scenarios" to solve sentence completion tasks from standardized reading exams. By modeling human activities as hierarchical events with temporal phases and probabilistic weights, the system performs abductive reasoning to identify the most likely situational "fillers" for missing text.

Introduction

Reading comprehension often requires more than linguistic parsing; it requires a model of the world. Consider a sentence from a primary school test: "Bob and Betty were happy as they piled bags of good things to eat in the car. They were going on a ____." While a computer can identify the parts of speech, it lacks the innate knowledge that "bags of food" and "car trips" are hallmarks of a picnic.

Candace Bullwinkle’s research addresses this "common sense" gap by introducing "scenarios"—data structures that represent the underlying framework of human activities. This approach moves beyond simple keyword matching, instead treating sentence completion as a problem of situational inference.

The Scenario Framework

The core of the system is a hierarchical organization of knowledge. Activities are not isolated facts but are nested within an "is-a-kind-of" tree. For instance, a "field trip" and a "picnic" are both categorized as types of "outings," which in turn fall under the broader category of "going to a place."

A hierarchical representation of the 'go X place' scenario, illustrating the 'is-a-kind-of' relationships used for classification.

This hierarchy allows the system to inherit properties from parent nodes while maintaining specific details at the leaf nodes. When the system encounters a sentence, it attempts to map the described actions to the most specific node possible in this tree.

Temporal and Probabilistic Logic

To accurately model events, the system divides each scenario into a timeline consisting of three distinct phases:

  1. Initiation: The preparatory actions (e.g., packing a basket).
  2. Performance: The central activity (e.g., eating at the park).
  3. Completion: The concluding actions (e.g., driving home).

Distinguishing between these phases prevents the system from confusing the preparation for an event with the event itself. Furthermore, the system employs a dual-weighting mechanism inspired by medical diagnostic models. Each observation is assigned an Upward Weight (how strongly an observation suggests a scenario) and a Downward Weight (how frequently an action occurs once a scenario is established).

The initiation phase of a scenario, showing the dual-weighting system (a and b weights) associated with specific observations.

As shown in the technical implementation, observations like "containing baskets/bags of food" are linked to specific weights (a1, b1) that help the system calculate the probability of a "picnic" scenario during the initiation phase.

Abductive Reasoning and Partitioning

The system processes text by generating an "Observations List" (OLIST). It then performs "dominance partitioning" to find the scenario that best explains the facts. If one scenario explains a subset of facts already covered by a more comprehensive scenario, the system prunes the redundant interpretation. This allows for "embedded scenarios"—for example, recognizing that a "grocery shopping trip" might be a sub-component of the "initiation" phase for a larger "picnic" event.

Evaluation and Results

The system was tested using examples from the Stanford Achievement Test and Gates Basic Reading Exams.

  • Successes: In one example, the system correctly identified "picnic" as the destination by linking "grocery shopping" facts to the broader outing scenario. In another, it successfully inferred "wig" and "library" by modeling an actress's behavior within a borrowing scenario.
  • Failures: The system struggled with a story about a cat and kittens. The failure occurred because the causal chain—linking a character being scratched to the protective nature of a mother cat—was too deep for the system's current one-level-deep inference limit.

Conclusion and Future Directions

Bullwinkle’s work demonstrated that situational context is a powerful tool for NLP. However, the research also highlighted significant hurdles. The numerical weights used for reasoning currently require manual tuning, and the system lacks the recursive depth needed for complex causal narratives. Future work must focus on automating the conversion of raw English into structured observations and expanding the temporal logic to handle varying verb tenses more fluidly. Despite these limitations, the scenario-based approach remains a foundational concept in the quest to give machines "common sense."

Summary updated

BibTeX
@inproceedings{bullwinkle1975picnics,
  author = {C. Bullwinkle},
  title = {{Picnics, Kittens And Wigs: Using Scenarios For The Sentence Competition Task}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1975},
}
Figure