The Knowledge Alignment Problem: Bridging Human and External Knowledge for Large Language Models