TL;DR
This paper proposes a computational model, the Tolerance Principle (TP), to explain how children learn language with remarkable uniformity despite varying input. It shows that the distribution of regular and irregular linguistic forms in a child's input, not just their presence, significantly influences learning speed and the likelihood of overgeneralization errors.
What Problem Does This Paper Address?
Child language acquisition is a fascinating paradox: children in diverse environments learn language with striking consistency, acquiring the same grammatical rules and following similar developmental paths. This uniformity persists even though the specific language input each child receives can vary significantly. The challenge is to explain how this consistent outcome arises from such variable input. Previous work has focused on general patterns, but this paper specifically investigates the role of input distributions – how frequently different forms appear – in shaping learning trajectories and individual differences.
Methods
The paper leverages the Tolerance Principle (TP), a type-based model of language productivity. The TP posits that a learner decides whether to generalize a rule or memorize exceptions based on a threshold ($\theta_N$). This threshold is determined by the number of known types that follow the rule versus the number of exceptions.
The authors extend the TP by:
- Quantifying Input Distributions: They introduce a method to assess the impact of how regular and irregular items are distributed within the input. This involves considering both type frequency (how many distinct forms exist) and token frequency (how often each form appears).
- Modeling Learning Trajectories: They develop a "Tolerance Principle state space" to visualize the learning process. This space maps a learner's progress based on the number of regular and irregular items learned. By simulating paths through this space, they can predict the probability of generalization at different stages of vocabulary development.
- Analyzing Distributional Effects: They specifically examine how different distributions of irregulars (e.g., concentrated at the high-frequency end of the input vs. spread out) affect these predicted learning trajectories.
Datasets
The paper uses simulated data derived from the CHILDES database for its case study on English past tense learning. Frequencies of past tense forms were extracted from this corpus. No new benchmarks were introduced; rather, existing data was analyzed through the lens of the proposed model.
Key Insights and Ideas
- Input Distribution Matters, Not Just Presence: The way irregular forms are distributed in the input (e.g., are they common or rare?) has a profound impact on learning, even if the number of irregulars is the same.
- "Top-Heavy" Irregulars Lead to Delayed Generalization: If irregular forms are mostly high-frequency, learners are less likely to generalize early on. They tend to learn these common irregulars first, delaying the point at which they might hypothesize a productive rule.
- "Bottom-Heavy" Irregulars Lead to Early Generalization: Conversely, if irregulars are mostly low-frequency, learners are more likely to generalize quickly. They encounter fewer exceptions early in their learning, making generalization seem more economical.
- "Split" Distributions Create Transient Variation: When irregulars are spread across both high and low frequencies, learning trajectories can be more complex, potentially leading to periods of overgeneralization followed by correction.
- Uniformity Explained by Type-Based Thresholds: The TP's reliance on type-based thresholds (rather than specific lexical items) helps explain why learners converge on similar grammars, even with varied input.
Key Results
- English Past Tense -ed: The model predicts a highly uniform learning trajectory for the regular past tense -ed, with generalization occurring around 400-500 vocabulary items. This aligns with empirical observations of consistent acquisition.
- English Past Tense Sing-Sang: For the irregular sing-sang pattern, the model predicts a much lower likelihood of productive generalization, with learners settling on non-productivity by around 12 vocabulary items. This also matches empirical findings from wug-test studies.
- Distributional Impact: The paper demonstrates that different distributions of irregulars for the same pattern (e.g., N=90, E=18) can lead to drastically different learning curves, with some distributions showing almost no generalization until late in development, while others show immediate generalization.
Limitations and Future Work
- Simplification of Input: The model simplifies input by focusing on type and token frequencies of specific patterns, potentially overlooking other crucial aspects of linguistic input.
- Empirical Validation: While the model aligns with some empirical data, the authors acknowledge the need for broader case studies to test its predictions across a wider range of linguistic phenomena and languages.
- Comparison with Other Models: The paper raises the question of whether other generalization models would yield similar or different predictions and which model best fits empirical data.
- Future Directions: The authors suggest quantitatively re-evaluating input distributions as a key factor in explaining cross-linguistic differences in child language development. They also propose exploring how other generalization models compare to the TP.
Summary updated