Recurrent Neural Language Models as Probabilistic Finite-state Automata