ISCA Archive SWAP 2000
ISCA Archive SWAP 2000

Activation flow in models of spoken word recognition

Uli H. Frauenfelder, Alain Content

Although current models generally agree that spoken word recognition involves the activation of a set of lexical competitors and the selection of the target word from this activated set, they differ in their view of the exact mechanisms underlying the selection process. Indeed, there are differences in how inappropriate competitors are eliminated from the activated set: bottom-up inhibition (Cohort), lateral inhibition (TRACE) and both (Shortlist). Moreover, the models also disagree in their assumptions concerning the role of top-down feed back in this process: existence of top-down feedback (TRACE) and no feedback (Cohort and Shortlist). Our research tests these divergent views with simulation and experimental studies.

Two phoneme detection experiments examined how lexical candidates are activated and deactivated by partial phonological mismatches at onset and item internal positions. The first experiment on initial one-feature mismatches at word onset provided evidence for late lexical activation. This result was more consistent with Shortlist than TRACE simulation findings. In the second experiment the mismatching information arrived after some initial lexical activation. The results showed complete lexical deactivation by the mismatching information, a result consistent with bottom-up inhibition assumed in the Shortlist model and not with lateral inhibition. These results taken together with other findings in the literature suggest that the selection process is based on the combination of the two inhibitory mechanisms.

Proponents of interactive models claim that top-down feedback facilitates word recognition, especially in cases where the input is defective. However, this claim about facilitatory feedback effects has been tested exclusively on phoneme recognition but not on word recognition. We evaluated this latter issue in a series of TRACE simulations in which we manipulated: the amount of phonological mismatch, lexicon size, and top-down feedback. Simulations showed that TRACE's recognition performance for small phonological mismatches was relatively poor (about 50%). This results suggest that, contrary to what is widely accepted, TRACE does not reliably recognize words with small initial mismatches. Moreover, when the parameter controlling the top-down feedback from word to phoneme levels was turned off, the recognition rate for both original and mismatch stimuli improved considerably with the larger lexica. These results suggest a detrimental rather than a beneficial effect of top-down feedback in word recognition. This paradox as well of the problems of scaling and parameters will be addressed.

Cite as: Frauenfelder, U.H., Content, A. (2000) Activation flow in models of spoken word recognition. Proc. Spoken Word Access Processes (SWAP), 79-82

  author={Uli H. Frauenfelder and Alain Content},
  title={{Activation flow in models of spoken word recognition}},
  booktitle={Proc. Spoken Word Access Processes (SWAP)},