Are There Preference Trade-offs in Attachment Decisions?

Lenhart K. SchubertAAAI 1986

Introduction: The Attachment Problem

In the field of computational linguistics, the "attachment problem" remains a central challenge in understanding how humans parse language. Consider the classic ambiguity: “John saw the bird with the binoculars.” Does the prepositional phrase "with the binoculars" attach to the verb "saw" (instrumental use) or the noun "bird" (the bird had binoculars)?

Historically, researchers split into two camps: those favoring rigid syntactic rules (e.g., "attach as low as possible") and those favoring "Preference Semantics," where word meanings and world knowledge drive the decision. Lenhart Schubert (1986) proposed a middle ground: a weighted potential model. He argued that human parsing is not a serial, rule-based flowchart but a dynamic system where syntactic and semantic factors compete using numerical weights.

The Weighted Potential Model

Schubert’s model calculates a "global potential" for various possible sentence structures. Instead of discarding options immediately, the parser evaluates them in parallel based on several competing factors:

  1. Syntactic Potentials:
    • Graded Distance Effect: Phrases generally prefer to attach to the nearest possible "head" word. However, larger, more complex phrases are more "mobile" and can tolerate greater distances.
  2. Rule Habituation: Familiar grammatical structures are "cheaper" for the brain to process than rare or complex ones.
  3. Semantic and Pragmatic Potentials:
    • Contextual Salience: If an object is already the focus of conversation, it is a more likely candidate for attachment.
  4. Script/Frame Conformity: Does the resulting meaning align with our understanding of the world? (e.g., birds rarely carry binoculars, but they often have feathers).

Experimental Evidence: The Linguistic Quiz

To validate this model, Schubert conducted psycholinguistic experiments using a "Linguistic Quiz." He presented subjects with carefully constructed sentence pairs and recorded their initial interpretations and "double-takes" (garden-path moments).

Subject response frequencies for various attachment scenarios.

As shown in Figure 1, the data reveals significant variability in how subjects resolve ambiguities. For instance, in sentence (16a), "The women discussed the children in the kitchen," 23 out of 29 subjects favored a specific attachment (likely the "low" attachment where the children are in the kitchen). However, in (16b), "The women discussed the dogs on the beach," that preference drops to 15/29. This shift demonstrates that changing the nouns (children vs. dogs) and locations (kitchen vs. beach) alters the semantic "potential," competing with the syntactic preference for proximity.

Key Insights: The "Huh? Oh!" Experience

One of Schubert’s most compelling arguments involves "garden-path" sentences. In the sentence, "Mary moved in with her uncle in New York who had fallen ill while on maternity leave," a staggering 86% (25/29) of subjects reported an initial anomalous reading. They initially attached "on maternity leave" to "uncle" because of the strong syntactic rule of "Right Association," even though it creates a biological impossibility.

This "Huh? Oh!" moment is crucial evidence for parallel competition. It suggests that syntax has a powerful, immediate potential that can temporarily override common sense. A purely serial semantic model would never make this mistake; a purely syntactic model would never correct it. Schubert’s weighted model explains both: the high syntactic weight causes the initial error, but the resulting semantic "penalty" eventually forces a re-evaluation.

Conclusion and Limitations

Schubert’s work suggests that serial rules are too "brittle" and "premature." By allowing multiple factors to compete in parallel, his model better reflects the flexibility of human cognition. However, the model leaves open the question of how these numerical weights are learned or tuned. Furthermore, while parallel evaluation is theoretically elegant, the computational complexity of maintaining multiple structural trees was a significant hurdle in the mid-80s. Future work remains to refine the "distance effect" and map exactly how constituent size influences these trade-offs.

Summary updated

BibTeX
@inproceedings{schubert1986are,
  author = {Lenhart K. Schubert},
  title = {{Are There Preference Trade-offs in Attachment Decisions?}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1986},
}
Figure