Title:
Hypothesis generation as a bridge between Human and Machine CogSci
Abstract:
The success of LMs in demonstrating non-trivially interesting linguistic behavior has raised a lot of excitement about the transfer of insights to (human) language science. But in order for this excitement to materialize, appropriate bridges need to be built from studies on computational models (Machine Cognitive Science) and those in humans (Human Cognitive Science). In this talk, I will discuss one possible bridge between the two systems: the development of hypothesis generation methods from which we can obtain novel predictions about human language that are sufficiently specific to be tested in the lab. I will present a case study where we used LMs trained on child-directed speech to generate novel hypotheses about the generalization of a novel dative verb to the alternate construction, given exposure to only one dative construction. That is, if pilked is observed in the double-object (She pilked me the ball), then under what conditions is it also found acceptable in the prepositional-object (He pilked a book to her)? Our method yields two new experiments that follow from our generated hypotheses, which we propose to test with human learners in the lab.