Investigating the Pre-Training Bias in Low-Resource Abstractive Summarization