|Implementing Escudero's model for the SUBSET problem
|This paper reports on the results of a series of experiments that put Escudero's theoretical model to a practical test. This model was developed to explain how Dutch learners of Spanish vowels solve the SUBSET problem. Escudero’s model uses Stochastic Optimality Theory and, in this paper, we use this same framework with the Gradual Learning Algorithm to simulate learning on the basis of Escudero’s proposal.
Escudero's model is based on empirical data which shows that Dutch learners initially use three native categories to classify the two Spanish vowels /i/ and /e/. The model proposes that this initial situation leads to non-optimal perception as well as non-optimal recognition because three words instead of two will be stored in this learner’s L2 lexicon. The empirical data also shows that advanced learners can attain optimal L2 perception. To explain this L2 development, the model proposes that both perceptual learning and lexical feedback are involved in the reduction of one of the three native categories to acquire optimal L2 sound perception and word recognition.
The proposal for the initial state in Dutch learners of Spanish vowels relies on the existence of minimal pairs in Spanish which
are assumed to lead to the storage of lexical entries different in form but identical in meaning and entries identical in form but different in meaning. Escudero proposes that the existence of two entries with the same form but with different meanings leads to what she calls semantic errors. The proposal also says that, when the learner notices these errors through contextual cues, perceptual learning will occur, just like in the lexicon-driven learning proposed in Boersma for L1 acquisition and in Escudero and Boersma for L2 acquisition.
In order to computationally test Escudero's proposal, several different parameter settings and correct candidate selection mechanisms were compared. It was found that the model as proposed in principle is viable but is very sensitive to noise. A different selection mechanism and alternate rankings remedy this problem, leading to complete category elimination and boundary shifts to the native positions.