Learning Stress Patterns with a Sequence-to-Sequence Neural Network

Brandon Prickett, Joe Pater

Abstract: We present the first application of modern neural networks to the well studied task of learning word stress systems. We tested our adaptation of a sequence-to-sequence network on the Tesar and Smolensky test set of 124 “languages”, showing that it acquires generalizable representations of stress patterns in a very high proportion of runs. We also show that it learns restricted lexically conditioned patterns, known as stress windows. The ability of this model to acquire lexical idiosyncracies, which are very common in natural language systems, sets it apart from past, non-neural models tested on the Tesar and Smolensky data set.