SFU Applied and Computational Math Seminar Series: Claire Moore-Cantwell

The abstract grammatical system of a language both shapes and is shaped by the sounds of the particular words that speakers know. I present GLaPL, a computational model which concurrently learns gradient grammar and gradient word representations. Features of a word (e.g. stress pattern) are represented with continuously-valued weights, simulating the memory strength associated with each feature of a word. These features compete with grammatical constraints to determine the realization of the word. I test this model on two data sets: corpus data of English comparatives ('more happy' v. 'happier'), and toy data representing a typology of stress systems with varying degrees of exceptionality. Under psychologically plausible assumptions about the learning process (more exposure to frequent words, memory decay for less frequent words), frequent lexical items emerge as more idiosyncratic than infrequent items. On the other hand, when a feature is completely predictable from the grammar, the model will fail to store it on individual lexical items. Small numbers of exceptions lead to intermediate ability to store a feature. These results are consistent with experimental and corpus evidence.