한국언어과학회

This study attempted to link continuous and dynamic articulatory information to categorical phonological feature representations through a neural network model. The Long Short-Term Memory (LSTM) model was used in the current paper, which is a type of recurrent neural network including temporal information connections. The test results of the LSTM model mapping muscular activation into phonological features show that (i) gradient values of phonological features are derived from the degree of activation of the tongue muscles, which determines the movement and shape of the tongue, and (2) the LSTM model can systematically capture vowels' co-articulatory effect on consonants.