Inherent Biases of Recurrent Neural Networks for Phonological Assimilation and Dissimilation

TL;DR

The paper investigates whether a lightweight recurrent neural network (RNN) can reproduce human-like biases in phonological pattern learning without resorting to alpha features or explicit constraint representations. Using a simple RNN with a single recurrent layer, inputs consisting of four phonological features across four segments are mapped to two classes (IN vs OUT) via a log-softmax output, trained online with gradient descent. Six patterns defined over consonant and vowel features reveal biases toward intra-dimensional, same-segment, and consonant-/vowel-only patterns, which the RNN captures without extra feature representations. When trained on full patterns and on subsets, the model demonstrates generalization and learning-time advantages consistent with human data, suggesting that sequential processing in RNNs can give rise to these biases. The work highlights the potential of recurrence to model phonological learning phenomena without explicit repeated-feature representations and motivates future exploration of unsupervised learning and scalable pattern complexity.

Abstract

A recurrent neural network model of phonological pattern learning is proposed. The model is a relatively simple neural network with one recurrent layer, and displays biases in learning that mimic observed biases in human learning. Single-feature patterns are learned faster than two-feature patterns, and vowel or consonant-only patterns are learned faster than patterns involving vowels and consonants, mimicking the results of laboratory learning experiments. In non-recurrent models, capturing these biases requires the use of alpha features or some other representation of repeated features, but with a recurrent neural network, these elaborations are not necessary.