Generating novel experimental hypotheses from language models: A case study on cross-dative generalization

Speaker: Najoung Kim Title: Generating novel experimental hypotheses from language models: A case study on cross-dative generalization Abstract: With the rapid advancement of neural network-based language technologies, how to make these models useful for the cognitive science of human language has been a topic of active debate. One idea that is often floated is to use these models as simulated learners to generate novel hypotheses that can in turn be tested with human experiments. I will discuss a case study in this direction and a set of methodology we developed for this goal, aiming for a more precise characterization of the exposure conditions that contribute to either licensing or blocking alternation in the human acquisition of novel dative verbs. From this case study, we derive a novel hypothesis that the extent to which generalization to the alternate form is facilitated depends on the harmonic alignment between features of the exposure context---in particular in the first postverbal argument. Based on these findings, we sketch out future experiments that can test this hypothesis in child learners.