Frame Selection in Parsing

Steven L. LytinenAAAI 1984

Introduction

In early Natural Language Processing (NLP), representing complex events relied on "frames" or "scripts"—predefined conceptual structures like $ARREST` or `$RESTAURANT. A persistent challenge in this paradigm was frame selection: the mechanism by which a system identifies the correct frame for a given text, especially when the language is ambiguous or vague. For instance, while the word "arrest" directly triggers an $ARREST frame, a phrase like "The police took him into custody" lacks a single definitive lexical trigger.

Historically, researchers attempted to solve this through massive sets of lexical "if-then" rules, frame-based expectations, or fixed decision trees known as discrimination nets. However, these methods often suffered from "rule explosion" or lacked the flexibility to handle initial frame selection in a scalable way. This paper introduces a more elegant approach: a refinement process driven by general inference rules traversing a conceptual hierarchy.

The MOPTRANS System: Hierarchical Refinement

The proposed system, MOPTRANS, shifts the focus from selecting a frame from a flat list to refining a vague concept into a specific one. This is achieved through two core components: a hierarchical conceptual memory and a set of general inference rules.

Hierarchical Conceptual Memory

Instead of isolated frames, concepts are organized into a hierarchy ranging from the abstract (e.g., ACTION) to the highly specific (e.g., POLICE-INVESTIGATION), linked by "IS-A" relationships. This structure allows the system to map vague words to high-level nodes and then "push" them down the tree as context becomes available.

Hierarchical conceptual memory for physical transfers (PTRANS) leading to specific events like earthquakes.

General Inference Rules

MOPTRANS replaces word-specific rules with three primary inference mechanisms:

  1. Script Activation: Mentioning one part of a stereotypical sequence activates the entire script and sets expectations for subsequent events.
  2. Expected Event Specialization: Vague words are automatically refined if they can be interpreted as abstractions of expected specific events within an active script.
  3. Slot-Filler Specialization: If a general action involves a prototypical actor (e.g., "Police"), the system searches the hierarchy for specific actions where that entity is the prototypical actor.

Case Study: Disambiguating "Hacer Diligencias"

The effectiveness of this hierarchical approach is demonstrated through the analysis of the Spanish phrase "hacer diligencias" (literally "to do diligent actions"). This phrase is extremely vague and its meaning depends entirely on context. In a grocery context, the system correctly refines this to SHOPPING.

In a police context—such as "The police are undertaking intense diligencias to capture a criminal"—the system utilizes the Slot-Filler Rule (recognizing police as prototypical actors of investigations) and the Expected Event Rule (recognizing that investigations typically precede captures).

Action hierarchy illustrating the relationship between general actions and specific scripts like police investigations and arrests.

By applying these rules to the hierarchy shown above, the system correctly refines the vague "diligencias" into the specific POLICE-INVESTIGATION frame. Crucially, the system achieves this using the same three general rules, rather than needing separate, hard-coded rules for every possible linguistic variation.

Conclusion

The MOPTRANS approach demonstrates that frame selection is best viewed as a refinement process rather than a simple choice. By leveraging a hierarchical organization of knowledge and a small set of general reasoning principles, the system can handle linguistic vagueness with greater efficiency and scalability than previous lexically-bound methods. While the system currently relies on manually engineered hierarchies, it provides a robust framework for language-independent conceptual parsing and contextual priming.

Summary updated

BibTeX
@inproceedings{lytinen1984frame,
  author = {Steven L. Lytinen},
  title = {{Frame Selection in Parsing}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1984},
}
Figure