Leveraging Unstructured Statistical Knowledge in a Probabilistic Language of Thought