Having Your Cake and Eating It Too: Autonomy and Interaction in a Model of Sentence Processing

Kurt P. Eiselt, Kavi Mahesh, Jennifer K. HolbrookAAAI 1993

Introduction

For decades, cognitive science and linguistics have been divided by a fundamental disagreement regarding how the human brain processes language. On one side, the Modular Camp (influenced by Fodorian theory) argues that syntax and semantics are handled by separate, autonomous processors. This view explains why humans are often susceptible to "garden-path" sentences, where we are led into a grammatical dead-end regardless of the sentence's potential meaning. On the other side, the Integrated Camp proposes a monolithic, connectionist process where all types of knowledge are blended from the start, explaining how semantic context can help us resolve grammatical ambiguities before they become problematic.

Experimental data, however, supports both positions. Humans exhibit modular behavior in some contexts and integrated behavior in others. Researchers at Georgia Tech and Albion College proposed a solution to this impasse: COMPERE (Cognitive Model of Parsing and Error Recovery). Their hybrid approach suggests that we can have a single processing engine that keeps its "knowledge books" separate, effectively allowing for both autonomy and interaction.

The COMPERE Model: A Unified Architecture

The core innovation of COMPERE is its architecture, which decouples the processing mechanism from the knowledge it uses. The model is built upon three primary pillars: a unified process, separate knowledge sources, and a uniform representation format.

The COMPERE architecture featuring a unified process interacting with separate knowledge sources.

As shown in the architecture diagram, a single Unified Process algorithm handles both syntactic and semantic information in tandem. While the process is unified, the data is not; syntactic rules, semantic roles, and conceptual meanings are stored in distinct networks. This allows the model to maintain the independence of different types of linguistic knowledge while ensuring they can influence each other during the real-time construction of a sentence's meaning.

Uniform Representation Across Domains

To enable a single engine to interpret two fundamentally different types of data (grammar and meaning), COMPERE utilizes a uniform representation. Both syntax and semantics are represented using a similar structure of nodes and links, typically organized as slot-filler pairs.

Uniform representation of lexical, syntactic, semantic, and conceptual information for the word 'moved'.

When a word like "moved" is encountered, the lexicon provides a lexical entry that triggers the creation of corresponding nodes in the syntactic, semantic, and conceptual networks. As illustrated above, the unified process can then rank feasible "bindings" or connections between these nodes. This bottom-up, early-commitment approach is guided by top-down expectations, allowing the model to build a coherent interpretation word-by-word.

Handling Linguistic Stress Tests

COMPERE was evaluated using controlled linguistic "stress tests" involving complex ambiguous sentences. A primary example is: "The bugs moved into the new lounge were found quickly."

This sentence presents two major challenges:

  1. Lexical Ambiguity: "Bugs" can refer to insects or electronic listening devices.
  2. Structural Ambiguity: "Moved" can be interpreted as the main verb (e.g., "The bugs moved themselves...") or the start of a reduced relative clause (e.g., "The bugs [that were] moved...").

In a "null context" where no prior information is given, COMPERE mimics human behavior by choosing the simplest grammatical path (the main verb interpretation). When it reaches the word "were," it realizes it has been "garden-pathed." However, because COMPERE treats error recovery as a core function, it retains alternative interpretations in the background. It can "unresolve" the ambiguity and switch to the relative clause path without needing to re-process the entire sentence from scratch.

Key Insights: Emergent Modularity

The most significant discovery of the COMPERE model is the concept of emergent modularity. The researchers demonstrated that "modular" behavior—such as being fooled by a garden-path sentence—does not necessarily prove the existence of a separate, isolated syntactic processor. Instead, it can be the natural result of a unified processor that simply has no semantic reason to prefer one path over another, thus falling back on a "syntactic default."

Furthermore, this model offers a new perspective on language impairments like aphasia. Rather than damage to a specific "processor," these conditions might result from damage to specific knowledge sources or the communication pathways between those sources and the unified engine.

Conclusion

By combining a unified processing algorithm with separate knowledge structures, COMPERE provides a computationally feasible bridge between the modular and integrated views of language. While the original prototype relied on hand-coded knowledge networks that are small by modern standards, the architectural principle—that processing can be unified even when knowledge is partitioned—remains a powerful insight into the complexity of human sentence processing.

Summary updated

BibTeX
@inproceedings{eiselt1993having,
  author = {Kurt P. Eiselt and Kavi Mahesh and Jennifer K. Holbrook},
  title = {{Having Your Cake and Eating It Too: Autonomy and Interaction in a Model of Sentence Processing}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1993},
}

Loading related content...

Figure