Introduction
In his 1984 AAAI paper, Garrison W. Cottrell introduced a connectionist model to address a long-standing puzzle in psycholinguistics: how the human brain navigates the "search problem" of words with multiple meanings. By utilizing a hierarchy of mutually inhibiting nodes, Cottrell’s model reconciles conflicting psychological data, demonstrating that while all meanings of a word may be initially accessed, strong context can suppress irrelevant meanings so rapidly that the process appears selective.
The Lexical Access Debate
In the early 1980s, cognitive scientists were divided into two primary camps regarding word recognition:
- Modularists: Argued that word recognition is a "bottom-up" process. Hearing a word like "bank" automatically triggers the retrieval of all its meanings (e.g., river bank, financial institution), regardless of sentence context.
- Interactivists: Argued that context (e.g., "I went to the river...") allows for "top-down" selection of only the relevant meaning immediately.
Research by Swinney, Tanenhaus, and others (STLB) identified an anomaly: while people generally access all meanings initially (supporting the Modular view), highly related "priming" words (like "ship" before "deck") seemed to lead to selective access of only the relevant meaning. Cottrell aimed to build a computational model explaining this "selective access" anomaly within a modular framework.
Methods: The Connectionist Model
Cottrell’s model is a precursor to modern deep learning, consisting of processing units (nodes) connected by weighted links. The architecture is defined by several key features:
- Discrimination Networks: The lexicon is organized hierarchically rather than as a flat list. A word node activates syntactic nodes (Noun vs. Verb), which in turn activate specific meaning nodes.
- Mutual Inhibition: Nodes at the same level compete. For instance, a strengthening "Noun" node actively suppresses the "Verb" node.
- Thresholding: Units only "fire" and send output once they accumulate sufficient evidence to cross a specific threshold.
- Weighted Feedback: Top-down feedback from context is allowed but weighted lower than bottom-up inputs. This enables context to "prime" a node—helping it reach its threshold faster—without causing the model to "hallucinate" words based solely on expectation.
The model was tested using a simulated lexicon centered on the word "deck," which possesses four distinct meanings: two nouns (SHIP-FLOOR, CARD-DECK) and two verbs (DECORATE, KNOCK-DOWN).
Key Results and the "Race" Condition
The model successfully replicated the STLB priming study results. It demonstrated that "multiple access" is the default state, but in strongly primed contexts, the "correct" meaning gains such a significant head start that it suppresses competitors before they can be detected in behavioral experiments.
In the simulation of "deck" with "ship" as a prime, the SHIP-FLOOR node reached its firing threshold at step 14. By step 24, it had successfully suppressed the CARD-DECK node. The model showed "selective access" for noun-noun ambiguities under strong priming but "multiple access" for noun-verb ambiguities. This occurs because competition between different syntactic classes (Noun vs. Verb) takes longer to resolve than competition within the same class.
Conclusion
Cottrell’s work highlights the importance of looking at the mechanism of lexical access—how nodes interact—rather than just the function. By predicting that "verb-verb" ambiguities would behave similarly to "noun-noun" ones, the model provided testable hypotheses for future research. While limited to low-level lexical access, this foundational connectionist approach paved the way for more complex models of sentence comprehension and high-level pragmatic processing.
Summary updated