Where is the answer? Investigating Positional Bias in Language Model Knowledge Extraction