Competitor effects in auditory word recognition: implications for interactive-activation models of word recognition

The TRACE model· of word recognition predicts the activation of "words within words" (tar in guitar). Simulations with TRACE are reported which demoostrate that the Ievel of this activation is affected both by the size of the set of competing word hypotheses (a "prefixing effect") and by the presence of a high-frequency competitor. A cross-modal priming study is described which supports and extends these predictions. INTRODUCTION: THE TRACE MODEL TRACE [1] is currently the most computationally explicit model of word recognition. It has an interactiveactivation architecture with three different Ievels of representation features, phonemes and words; the parameters governing the interactions within and between these Ievels may be independently manipulated. The very existence of these different types of interaction in the human speech processing mechanism (HSPM) is a fundamental theoretical issue. TRACE's authors claim empirical support for these interactions from perceptual studies. When TRACE receives an input, numerous lexical hypotheses are activated, which then compete with each other until one predominates. TRACE makes a number of predictions about this competition which have testable perceptual implications. In this paper, TRACE's basic mechanism is described, then a number of predictions are demonstrated, and finally a cross-modal priming study [2], measuring transient word-activations, is described which offers experimental support for TRACE's behaviour. Input to TRACE consists of a sequence of feature-level descriptions made available across a number of discrete time-slices. TRACE's architecture is localist: an individual node stands for a specific feature, phoneme or word at a specific point in time. There are separate nodes representing the phoneme /gl, for instance, which monitor the presence of that phoneme across overlapping sequences of six time-slices. If the activation of one of these nodes is high, then the model is hypothesizing the phoneme lgl at the point in time monitared by that node. There is intra-level inhibition, and inter-level activation: activated nodes within. any Ievel compete to suppress each other, while mutally consistent nodes in adjacent Ievels reinforce each others activation Ievel. As the feature-level nodes become activated by the input, they activate the relevant phoneme-level nodes, which in turn generate activation at the word-level. Activation of nodes within a Ievel allows competition to occur within that Ievel. Activation at the word Ievel allows the topdown activation of consistent nodes at the phoneme Ievel. A complex pattern of acti vation results, reflecting TRACE's ability to allow all of the words in its lexicon to compete at each point in time to the extent that they are activated by the input. TRACE's output is an activation Ievel for each node for each point in time. Plotting these Ievels against time produces the activation curve for a particular lexical or phonemic hypothesis. The actual recognition of a word is an independent issue which may be modelled by a decision rule which compares the relative activations of the word nodes. PREDICTIONS FROM TRACE TRACE makes the following predictions. (i) A sequence of segments which is homophonaus with a monosyllable, but which is embedded within a Ionger word will, in principle, be able to activate the representation ofthat monosyllable. That is, the second syllable of the input guitar should cause the activation of tar. (Henceforth, "the activation of :r." will be used to mean "the activation of the node representing :r..") Figure 1 shows the activation curve for guitar when the input is a feature-level description of that word; there is also a brief activation of tar caused by the latter half of the input. This happens because segmentation in TRACE does not occur prior to lexical access/activation. All activated words compete, even though they may only partially overlap with the input