Phonotactic learning with neural language models

Computational models of phonotactics share much in common with language models, which assign probabilities to sequences of words. While state of the art language models are implemented using neural networks, phonotactic models have not followed suit. We present several neural models of phonotactics and compare their performance to a commonly employed phonotactic model. We show that these models are better able to learn longdistance dependencies, do not require stipulation of a feature system, and agree more closely with human judgements. This work provides a promising starting point for future modeling of human phonotactic knowledge.