Proceedings of the Society for Computation in Linguistics

A neural-based gradient ascent algorithm for learning weights and activations that correctly predict output forms, by optimizing the Harmony of a whole paradigm by exploiting a blurring of the distinction between stems and affixes.