Computational Models of Generalization and Item Specificity

Course Description

A central tension in linguistics is how to account for the fact that languages contain generalizable structure—which makes linguistic productivity possible—but also rampant idiosyncrasies in how individual words and phrases are used. For example, an English speaker knows that the past tense of a novel verb glorp is glorped but the past tense of runs the irregular ran. This item-specific knowledge is not limited to strict irregularities, but also includes knowledge of quasi-regularities (e.g. Feel-felt, deal-delt, etc.) and statistical usage preferences such as verb subcategorization preferences and frequent collocations. Despite decades of debate, linguists still disagree about how these two types of knowledge relate to each other and to what extent each is recruited in language acquisition and processing. Computational modeling provides a new path to address this old question, allowing us to formalize and make testable predictions about the joint roles of generalization and item-specificity in language acquisition and processing. Advances in computing power, the development of new statistical methods, and the creation of large linguistic datasets are all contributing to new theoretical advances on this front. In this course, we will focus on computational models of how generalization and item-specific knowledge jointly contribute to language acquisition and processing. We will introduce different classes of models, focusing on 1) connectionist and modern neural network models; 2) Bayesian and information-theoretic models; and 3) exemplar models. We will apply these models to case studies in syntax and semantics, in both acquisition and processing.

Area Tags: Psycholinguistics, Computational Linguistics, Syntax, Semantics, Acquisition, Statistics

Instructors: Emily Morgan & Masoud Jasbi