Abstract
While inflectional morphology is broadly rule-governed, many inflectional paradigms admit exceptions (e.g., the past tense of "go" is not "goed" but "went"). Regularity in form-meaning mapping permits generalization and facilitates learning, and regularity has been shown to emerge through language transmission. However, this research demonstrates an "irregularization bias," where having high-frequency words be irregular prevents them from interfering with our ability to generalize regular patterns to new words. By "filtering" high-frequency items into storage, the cognitive system can more effectively identify productive rules across low-frequency items.
Introduction: The Tug-of-War Between Rules and Exceptions
In linguistics, we often see a tug-of-war between rules (regularity) and exceptions (irregularity). While rules make language efficient to learn and use, almost every language is peppered with irregular forms (e.g., eat/ate instead of eat/eated).
The traditional view is that irregulars are a "by-product" of history—high-frequency words are used so often that they resist being "corrected" by rules over generations. However, this paper asks a more provocative question: Does irregularity serve a functional purpose? Specifically, the authors investigate whether irregularity in high-frequency items actually makes the rest of the language's regular rules easier to learn by reducing "storage interference."
Methods and Experimental Design
The researchers conducted a behavioral experiment with 46 adult participants to test a hypothesis derived from a computational model by O’Donnell (2015). The core idea is a trade-off between computation (using a rule) and storage (memorizing a word).
Participants were tasked with learning an artificial language consisting of 48 novel inflected words divided into six "paradigms" (groups). The study used three conditions to manipulate the relationship between frequency and regularity:
- Fully Regular: All words in the group used the same suffix and appeared with equal frequency.
- Frequent Regular: One word appeared very frequently (48 times) and followed the regular rule, while others appeared rarely.
- Frequent Irregular: The most frequent word (48 times) used a unique, irregular suffix, while the low-frequency words followed a regular rule.
The training followed a Zipfian-like distribution (one very common word, many rare words) to mimic natural language. After training, participants were tested on their ability to recognize words they had seen and, crucially, their ability to generalize the regular rules to entirely new stems.
Key Insights: The Irregularization Bias
- The Storage Interference Effect: Our brains tend to store high-frequency words as whole units rather than calculating them using rules. If a very common word is regular, it "clogs" the system; the brain stores it anyway, which makes the underlying rule look less "productive" or useful for general application.
- Irregularity as a "Filter": When a high-frequency word is irregular, it is effectively "set aside" by the learner's inference engine. This allows the learner to see the regular rule as something that applies consistently across a wide variety of low-frequency items—a hallmark of a truly productive rule.
- Productivity Signaling: The brain infers how productive a rule is by seeing how many different words it applies to, especially low-frequency ones. High-frequency regulars actually inhibit this inference.
Results
- Better Generalization: Participants were significantly more likely to apply regular rules to new words if the high-frequency item in that category was irregular.
- The Regularity Penalty: In the "Frequent Regular" condition, the presence of a high-frequency regular word actually decreased the participants' willingness to use that same rule for new words.
- Storage Advantage: High-frequency irregulars were memorized and accepted more reliably than high-frequency regulars of the same frequency. This suggests that the "weirdness" of an irregular form makes it more "sticky" in memory when it occurs often.
Discussion and Future Directions
The authors acknowledge that while this lab experiment shows a learning bias in individuals, it doesn't yet prove how this plays out over centuries of cultural evolution.
Future directions include: * Iterated Learning Studies: Testing whether this bias causes irregulars to spontaneously emerge and stabilize when language is passed from "teacher" to "learner" over multiple generations. * Cross-Linguistic Analysis: Examining whether natural languages across the globe show the specific frequency-to-irregularity distributions predicted by this model. * Computational Refinement: Further exploring the "computation vs. storage" trade-off in more complex morphological systems beyond simple suffixes.
Summary updated