Refining attention for explainable and noise-robust fact-checking with transformers