Exploring Phonotactics with Simple Recurrent Networks
Ivelin Stoianov, John A. Nerbonne · 2000
Stoianov, Nerbonne and Bouma (1998) trained Simple Recurrent Networks (SRNs) on graphotactics of Dutch monosyllabic words, overcoming shortcomings of previous implementations. The current report is a continuation of our earlier research, but using phonetic data representations instead of orthographic, that is, learning phonotactics. In addition, we conducted further analysis of neural network performance with regard to some variables such as word frequency, length, neighbourhood density and error location. The results are compared with reported psycholinguistics analyses. This informal comparison of SRNs and human performance suggests that SRNs can be used for modeling natural language processing.