Predicting high-level human judgments with word embeddings

Every day, people judge lots of objects and entities in a variety of ways; we judge the tastiness of foods, the sense of significance a job might bring us, the masculinity of personality traits, and more. Despite the ubiquity of these judgments, psychologists do not currently have a general and effective approach to predicting and understanding these judgments, as it has been difficult to capture the object/concept knowledge that subserves these judgments. In this work, we show that by regressing judgments directly on word embeddings — high-dimensional vectors of word representations trained on large language corpora — we can predict out-of-sample judgments with a high degree of accuracy. This is work done with Wanling Zou and Sudeep Bhatia at the Computational Behavioral Science Lab at University of Pennsylvania. Paper coming soon!

Conventionalization of lexicons in social networks

In my master’s work, we showed (1) that Nicaraguan Sign Language conventionalized a basic lexicon faster than did four different homesign systems, and (2) a group of simulated agents in a social network like that of NSL (roughly, a fully connected one) conventionalized faster than a group of agents in a social network like that of homesign (a star network, with the homesigner at the middle). We thus have suggestive evidence that full networks are faster at conventionalization than are star networks. With Matt Hall and my former co-advisor Marie Coppola, we tested these network effects experimentally with hearing individuals brought into a lab setting. Paper coming soon!

Are prosodic representations amodal?: An ERP investigation

Which properties of languages and linguistic representations are shaped by the channels in which they are used – auditory-oral for spoken languages and visual-manual for signed languages — and which properties are shared across channels, suggesting that they reflect an amodal or supramodal language system? With a host of collaborators across multiple research institutions, we are investigating whether amodality might extend to processing of prosodic phrasing, by comparing ERP responses to intonational phrase boundaries in a signed language (ASL) to those in a spoken language (English). Update Fall 2018: results will be presented at Psychonomics and at ETAP this fall….