A Computational Theory of Processing Overload and Garden-Path Effects

Edward GibsonCOLING 1990

TL;DR

This paper proposes a computational theory of sentence processing that explains why certain sentences are difficult or impossible to understand (processing overload) and why we get "stuck" on certain sentence structures (garden-path effects). It posits that these phenomena arise from the limited capacity of working memory during parsing, where the "load" is determined by specific syntactic properties related to thematic roles and lexical requirements.

What Problem Does This Paper Address?

Human sentence processing is remarkably efficient, yet we all encounter sentences that are difficult to parse, leading to confusion or outright incomprehension. These difficulties can manifest as "garden-path effects," where we initially interpret a sentence in a way that turns out to be incorrect, forcing us to reanalyze it, or as general processing overload, where a sentence is simply too complex to process at all. Existing theories often struggle to provide a unified account for these phenomena, particularly how syntactic complexity directly translates into processing difficulty. This paper aims to bridge this gap by proposing a computational model that quantifies processing load based on specific linguistic principles.

Methods

The paper introduces a computational model of sentence parsing that assumes a parallel architecture, meaning multiple interpretations can be considered simultaneously. The core of the model lies in assigning a "processing load" (measured in Processing Load Units or PLUs) to different syntactic structures. This load is determined by two key principles derived from Government and Binding Theory:

  • Property of Thematic Reception (PTR): Assigns a load when a constituent is in a position to receive a thematic role (e.g., subject, object) but its thematic role assigner is not yet identifiable.
  • Property of Lexical Requirement (PLR): Assigns a load when a lexical requirement of a word (like a verb needing a complement) is not yet satisfied by a hypothesized structure.

The model proposes that a sentence becomes unacceptable due to overload when the cumulative processing load exceeds a maximum capacity (K). Furthermore, it suggests that preferred readings of ambiguous sentences and garden-path effects occur when the processing load difference between competing interpretations exceeds a certain preference threshold (P). The paper also introduces a parser architecture with a buffer and a stack-set to accommodate parallel processing.

Datasets

Not discussed in the paper. The paper relies on analyzing existing linguistic examples and constructing novel sentences to test its hypotheses.

Key Insights and Ideas

  • Quantifying Syntactic Load: The paper's central innovation is quantifying the processing load associated with specific syntactic properties (thematic role assignment and lexical requirements). This moves beyond qualitative descriptions of complexity to a measurable metric.
  • Unified Account of Overload and Garden-Paths: The model provides a single framework to explain both processing overload (unacceptability) and garden-path effects (misleading interpretations), linking them to the same underlying mechanism of working memory limitations.
  • Parallel Parsing with Constraints: By assuming a parallel parsing architecture, the model can account for the existence of multiple interpretations, but then uses the processing load to explain why only one is preferred or why some become problematic. This contrasts with purely serial models.
  • Cross-Linguistic Predictions: The theory is tested not only on English but also makes predictions for other languages (Dutch and German), suggesting that the proposed syntactic principles and their associated processing loads are universal.

Key Results

The paper demonstrates that the proposed model can successfully predict:

  • Garden-Path Effects: Sentences like "# I put the candy on the table in my mouth" and "# The horse raced past the barn fell" are predicted to cause garden-path effects due to significant processing load differences between competing interpretations.
  • Processing Overload: Center-embedded relative clauses, particularly those with more than two levels of embedding, are predicted to be unacceptable due to exceeding the maximum processing load (K). For instance, sentences with three initial noun phrases requiring thematic roles are predicted to be acceptable (load of 3xInt PLUs), while those with four are predicted to be unacceptable (load of 4xInt PLUs).
  • Preferred Readings: Ambiguous sentences like "John expected Mary to like Fred" are predicted to have both interpretations maintained if the load difference is below the preference threshold (P), explaining why they are not difficult to parse.
  • Cross-Linguistic Consistency: The model's predictions align with observed acceptability patterns in German and Dutch, particularly concerning center-embedded structures and verb-final subordinate clauses.

Limitations and Future Work

  • Parameter Values: The paper acknowledges that the specific values for the processing load units (XTR, XLR) and the preference threshold (P) and maximum capacity (K) are not definitively determined and require further empirical investigation. The assumption that XTR = XLR = XInt is made for consistency but could be refined.
  • Formalization of Full Interpretation: The authors note that the principles of Thematic Reception and Lexical Requirement are derived from the broader principle of Full Interpretation, which itself is not yet fully formalized. A more complete formalization of Full Interpretation could potentially unify these properties further.
  • Scope of Properties: While the paper focuses on the 0-Criterion and Projection Principle, it suggests that other syntactic principles could also contribute to processing load and might be incorporated into the theory.
  • Empirical Validation: While the paper uses existing examples and constructs new ones, more extensive experimental data would be needed to fully validate the quantitative predictions of the model.

Summary updated

BibTeX
@inproceedings{gibson1990computational,
  author = {Edward Gibson},
  title = {{A Computational Theory of Processing Overload and Garden-Path Effects}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1990},
}
Figure