Is There a One-Model-Fits-All Approach to Information Extraction? Revisiting Task Definition Biases