I will discuss a new model of phonological encoding, which is implemented as a simple recurrent connectionist network (Elman, 1990; Dell, Govindjee, & Juliano, 1993). Each word in a large vocabulary is represented by a static, lexically specific input pattern, and the model learns to map this static pattern onto a temporal sequence of phoneme specifying output patterns.
The recurrent feedback forms the basis of the model's ability to enforce serial order upon phonemes. At each point in time, the hidden layer receives a copy of its previous state via feedback connections. This causes the network's computations to be influenced by previous computations. Thus, at each point in the word, computations at earlier points provide contextual cues for the current phoneme.
As the model produces a word, the input pattern identifies a lexical item and its phonemes but not the order in which the phonemes occur or the number of occurrences of each phoneme. The input layer includes one phoneme unit for each phoneme in the language and also a set of units that provide a surrogate for semantic features. The input pattern remains activated for the duration of the word, which consists of one time step for each phoneme in the word. At each time step, the model's output is a pattern of activation indicating the phonological features of the current phoneme.
The model acquires a vocabulary of 1200 words, ranging in length from one to three syllables (three to twelve phonemes). After approximately 200 epochs of back-propagation training on the vocabulary, the model produces errors on fewer than 0.1% of the phonemes in the vocabulary. This lexicon is substantially more complex than that of most other models of phonological encoding, which is important in two respects. First, it provides a test of the learning and representational capacities of simple recurrent networks. At the same time, the large lexicon exposes the model to the distributional properties of a large sample of the language, allowing it to acquire knowledge of a wide range of phonological regularities, both fine- and course-grained. The impact of this structural knowledge is observed across the vocabulary.
The model provides a unified account of two prominent issues related to phonological encoding, which have often been treated independently by psycholinguistic theory:
Damaging the model through the addition of noise to the input layer results in output patterns that show several characteristics of normal and aphasic speech errors, which will be discussed.
References
Dell, Govindjee & Juliano (1993). Structure and content in language production: A theory of frame constraints in phonological speech errors. Cognitive Science, 17, 149-195.
Elman, J. (1990). Finding structure in time. Cognitive Science, 14, 179-211.