Achievement

A fundamental generalization about phonological grammars is that assimilation - a type of process found in natural language by which one sound becomes more similar to another - preferentially applies to sounds that are independently similar, either because they already share one or more properties or because they occupy similar positions within the word. The dissertation of IGERT trainee Adam Wayment develops a novel explanation for this generalization by integrating ideas from formal linguistics with ideas of neural network computation. This similarity-breeds-similarity effect is shown to follow from a novel type of connectionist architecture, *entailment networks*. At a higher level of representation, the computations and biases of entailment networks can be instantiated as families of *attraction constraints* within Optimality Theory. These new constraints enable a grounded symbolic account of a wide range of assimilation patterns found in the languages of the world.