Inductive Bias And Modular Design For Sample-Efficient Neural Language Learning