Causal Direction of Data Collection Matters: Implications of Causal and Anticausal Learning for NLP