Probing Neural Network Generalization using Default Patterns

Brandon Prickett, Tianyi Niu, Katya PertsovaNAACL 2025

TL;DR

This paper investigates how neural networks learn and generalize "default" linguistic patterns, particularly minority defaults, using a simplified artificial language. The study reveals that while neural networks can learn these patterns, frequency plays a dominant role, and distribution also influences learning speed, mirroring human learning tendencies.

What Problem Does This Paper Address?

The ability of neural networks to generalize linguistic rules, especially those that act as "defaults" (rules for a heterogeneous category that apply as a last resort), has been a long-standing question. Previous studies using complex natural language data have yielded ambiguous results. This paper aims to clarify this by using a controlled artificial language to isolate the learning of default patterns, focusing on how frequency and distribution affect generalization, particularly for minority defaults (where the default rule applies to a less frequent category).

Methods

The researchers employed three computational models to learn the artificial language pattern:

  • Encoder-Decoder (ED) RNN: A recurrent neural network architecture that processes sequences sequentially, acting as a form of memory.
  • Transformer Encoder: A model that utilizes self-attention mechanisms to process all input segments simultaneously, differing from the ED's sequential approach.
  • Logistic Regression (LR): A simpler, more interpretable model used for comparison, with two variants: one that considers all segments (LR-PC) and one that only considers the last segment (LR-PL).

The models were trained on an artificial language designed to mimic English plural allomorphy, where different suffixes were assigned based on the final segment of a word. The study manipulated the frequency of the "default" suffix across three conditions: Equal Frequency, Majority Default, and Minority Default.

Datasets

The study used a custom-generated dataset for training and evaluation. This dataset was derived from the CMU Pronouncing Dictionary and consisted of words generated using specific consonant-vowel (CV) templates. The dataset was structured into three conditions based on the frequency of the global default suffix (Suffix C):

  • Equal Frequency: All three suffixes (A, B, C) appeared with equal frequency.
  • Majority Default: Suffix C (the global default) appeared in 90% of the words.
  • Minority Default: Suffix C appeared in only 10% of the words.

The evaluation included various types of stimuli to test generalization:

  • Test data: Novel words with the same segments and templates as training data.
  • Mutants: Words where the final segment was changed to a different one, testing sensitivity to segmental similarity vs. relevant features.
  • Novel Templates: Words with CV-shapes not seen during training.
  • Novel Segments: Words ending in segments not present in the training data.

Key Insights and Ideas

  • Frequency is King: The most significant factor influencing learning speed across all models was the frequency of a pattern. Lower frequency patterns were consistently learned slower.
  • Distribution Matters, But Less So: While frequency dominated, the distribution of patterns also played a role. Patterns with a wider, more heterogeneous distribution (defaults) were learned slower than those with a narrower distribution, even when frequency was held constant.
  • Minority Defaults are Learnable, but Difficult: All models were capable of learning minority default patterns, but they were learned more slowly compared to other categories.
  • Neural Networks Mimic Human Learning: The observed effects of frequency and distribution on learning speed align with findings from human language acquisition and category learning studies.
  • Models Generalize, But with Caveats: The models showed good generalization to novel words and segments, but struggled with certain novel templates (e.g., ED with CVCC, Transformer with VC). This suggests potential limitations in how models represent concepts like "last segment" or their reliance on word shape.
  • Defaults as Competition: Model analysis revealed that default patterns are learned through a competition among different feature-based categories, where more specific classes can "block" more general ones.

Key Results

  • All three models (ED RNN, Transformer Encoder, Logistic Regression) achieved high accuracy on the test data, demonstrating their ability to learn the artificial language pattern.
  • In the Minority Default condition, models learned the default suffix (C) slower than other suffixes, but eventually achieved high accuracy.
  • The ED model performed perfectly on novel words with the same segments and templates as training data after about 25 epochs.
  • The Transformer Encoder showed a tendency to overgeneralize the default suffix (C) in the Majority Default condition during early learning stages.
  • Both neural network models performed perfectly on "mutant" words, indicating they relied on the relevant features of the last segment rather than overall word similarity.
  • The LR model's analysis provided interpretable insights into feature weights, showing how features like [voice] and [coronal] contributed to suffix assignment.

Limitations and Future Work

  • Artificial Language Simplicity: The artificial language, while useful for isolating specific phenomena, is simpler than natural language. The authors acknowledge that more complex scenarios, like exceptions not forming a neat natural class, would be a valuable next step.
  • Novel Template Issues: The ED model's random behavior on CVCC templates and the Transformer's struggles with VC templates suggest potential limitations in how these models handle novel word structures, particularly those with fewer segments than seen in training.
  • "Across-the-Board" Property of Defaults: The authors plan to investigate how models handle the "across-the-board" property of defaults, which some rule-based models predict should apply to novel situations like abbreviations or proper names. Testing with contrastive features not seen in training is a proposed direction.
  • Type vs. Token Frequency: The current study only varied type frequency. Future work will explore the impact of token frequency, which is known to play a different role in language acquisition.

Summary updated

BibTeX
@inproceedings{prickett2025probing,
  author = {Brandon Prickett and Tianyi Niu and Katya Pertsova},
  title = {{Probing Neural Network Generalization using Default Patterns}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
}
Figure