An Explanation for Minimal Attachment and Right Association

Robert MilneAAAI 1982

Abstract

In the study of human sentence processing, two principles have long been used to describe how readers resolve structural ambiguities: Minimal Attachment (MA) and Right Association (RA). While these principles accurately describe behavior, they have historically been treated as independent, stipulated rules within psycholinguistic models. This paper argues that MA and RA are not primary psychological rules but are instead emergent side effects of the architecture of deterministic production system parsers. By comparing the Augmented Transition Network (ATN) framework with the ROBIE parser, this work demonstrates that these parsing strategies are functional necessities of rule prioritization and local data-driven processing.

Introduction

When humans process language, we do not wait until the end of a sentence to begin building its structure. Instead, we resolve ambiguities incrementally. In the late 1970s, Lyn Frazier and Janet Fodor formalized this observation through two key principles:

  1. Right Association (RA): The preference to attach new constituents to the lowest, most recently processed part of the current structure. For example, in the sentence "Joe called the friend who smashed his car up," the particle "up" is naturally associated with "smashed" rather than "called."
  2. Minimal Attachment (MA): The preference for the simplest possible structure that incorporates new input with the fewest additional nodes. In "John bought the book for Susan," the phrase "for Susan" is attached to the verb phrase ("bought for Susan") rather than the noun phrase ("the book for Susan"), as the former requires fewer structural additions.

While models such as the "Sausage Machine" or Augmented Transition Networks (ATNs) could replicate these behaviors, they did so by manually scheduling specific operations. This paper seeks to move beyond mere description to provide a functional explanation for why the human parser operates this way.

Methodology: From ATNs to Production Systems

The author evaluates these linguistic principles by contrasting the traditional ATN approach with ROBIE, a deterministic parser based on Mitchell Marcus’s PARSIFAL.

In an ATN, the parser moves between states via "arcs" that represent different types of linguistic transitions. To simulate human behavior, researchers had to explicitly order these arcs—for instance, ensuring a "WORDarc" was checked before a "CAT arc." In contrast, ROBIE utilizes a production system grammar. Here, rules (productions) are triggered by specific patterns in the input buffer and the current node stack.

The core of this system is the "Most Constrained Pattern First" principle. Rules are prioritized based on their specificity:

  • Specific Patterns: Rules matching specific words or sequences are tried first.
  • General Categories: Rules matching grammatical categories (e.g., Noun, Verb) are tried next.
  • Default Rules: Rules with no specific pattern (often represented as [t]) are tried only if all others fail.
Example production rules in the ROBIE parser, demonstrating the hierarchy from specific word patterns to general categories.

This hierarchy ensures that the parser always attempts the most precise interpretation of the data before falling back on general structural assumptions.

Theoretical Framework: Emergent Principles

The central claim of the paper is that RA and MA are necessary consequences of this production-based architecture.

Right Association as a Default

In ROBIE, the act of "closing" a phrase—which corresponds to the behavior of Right Association—is handled by a default rule. Because the system always prioritizes more constrained rules that might attach a new word to the current open phrase, it will only "close" that phrase and move to a higher level if no specific attachment can be made. Thus, RA is not a rule the parser chooses to follow; it is the result of the parser exhausting all local attachment possibilities.

Minimal Attachment as Local Processing

Minimal Attachment emerges from the "data-driven" nature of the parser. ROBIE lacks "SEEK" arcs, which in an ATN proactively look for new, potentially complex phrases. Instead, ROBIE must use incoming words to satisfy the requirements of the current local structure. Because the parser cannot "look up" the stack to see if a word might fit into a more complex, higher-level structure without first failing at the local level, it naturally prefers the simplest, most immediate attachment.

Comparative Analysis

The paper provides a qualitative proof-of-concept by showing that the manual "scheduling" required in ATNs is inherent in the production system. In ATNs, achieving human-like parsing required scheduling "SEND" and "JUMP" arcs last.

Standard arc types in Augmented Transition Networks (ATNs) used to model linguistic transitions.

As shown in Figure 2, these arcs represent the termination of a node or optionality. In ROBIE, the equivalent of these arcs are the default rules. By placing these at the bottom of the priority list, ROBIE achieves the same effect as the ATN's manual scheduling, but as a fundamental property of the system's logic rather than an arbitrary setting.

Conclusion

By shifting the perspective from psychological rules to computational constraints, Milne provides a compelling explanation for the origins of Minimal Attachment and Right Association. These principles are revealed to be side effects of a parser optimized for deterministic efficiency. This suggests that the human linguistic faculty may be structured as a production system where the "Most Constrained Pattern First" logic serves as the primary engine of sentence comprehension.

Reference: Milne, R. W. (1982). An Explanation for Minimal Attachment and Right Association. AAAI-82 Proceedings.

Summary updated

BibTeX
@inproceedings{milne1982explanation,
  author = {Robert Milne},
  title = {{An Explanation for Minimal Attachment and Right Association}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1982},
}

Loading related content...

Figure