Sources of Hallucination by Large Language Models on Inference Tasks