Phenomena in spoken language processing that can be explained using representations
at one level of processing often have an alternative explanation based on distributional patterns of units at a lower level of analysis. The absence of (a) good corpora for spoken language and (b) good theories of the relevant distributional statistics makes it difficult to distinguish among these alternatives. Research with artificial languages offers a promising methodological tool, allowing for precise control over the properties of the input. After demonstrating that it is feasible to explore word recognition using an artificial lexicon, we use an artificial language to examine whether lexical feedback affects perception in compensatory co-articulation as proposed by Elman and McClelland (1988) or whether all compensatory effects can be explained by probabilistic phonotactics as proposed by Pitt and McQueen (1998).
We first present research by Magnuson and colleagues in which participants learned to associate names with novel shapes in a forced choice training paradigm with feedback. A 16 word artificial lexicon was constructed so each word had a cohort competitor (e.g., pibu had pibo) and a rhyme competitor (e.g., pibu had dibu) The frequency of presentation of the words and their competitors varied systematically during training. During testing, eye movements were measured as subjects followed the same type of instructions used in training. After less than two hours of training, subjects were processing the words incrementally, making eye movements to the target shape before the end of the word. We found frequency effects, cohort and rhyme effects, and neighborhood density effects that closely match those found with real words.
We then report ongoing research that is examining perceptual effects of compensatory co-articulation using a richer artificial language. The language includes eight nouns, each associated with a novel shape and eight adjectives, each associated with a different texture. The crucial stimuli are adjectives ending in s or esh followed by nouns beginning with k or t. The words are synthesized to preserve the acoustic pattern associated with compensatory co-articulation in natural English. The test stimuli are synthesized so that some of the adjectives end in an ambiguous s/esh and the nouns vary along a five-step t/k continuum. The participants task is to click on the appropriate shape, among four alternatives.
The language is constructed to allow for a clear test of whether lexical status has effects on compensatory co-articulation above and beyond effects that can be attributed to probabilistic phonotactics.