Text this: Analyzing Semantic Faithfulness of Language Models via Input Intervention on Question Answering