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