Causal relations play a major role in many tasks in Natural Language Understanding (NLU) (Girju, 2003) and discourse analysis (Mulder, 2008; Kuperberg et al., 2011). The relation between X and Y is considered causal when X makes Y happen or exist, or vice versa, where X and Y can be an event, state, or object. For example, in ”The river had now turned into full flood after the deluge of rain a few days ago.”, deluge of rain is the cause of flood which happens when there is an overflow of water.
Automatic extraction of causal relations is a challenging task in Natural Language Processing (NLP). Previous work on identifying causal relations is mainly focused on classifying pairs as causal/non-causal without necessarily considering direction in the pairs. These lines of research either assume a fixed direction between spans in a pair or do not directly test models in predicting the direction in a causal pair. For example, (Gao et al., 2019a; Liu et al., 2020) focus only on identifying causal pairs in context and not specifying the direction in the pairs, namely which entity caused the other. The knowledge-oriented CNN (K-CNN) model for causal relation extraction model (Li and Mao, 2019) attempts at finding sentences that include a causal relation only by feeding a pair of spans and a sentence as input to the model. On the other hand, (Dunietz et al., 2018) indirectly addresses the direction prediction problem by finding cause and effect spans associated with only explicit causal connectives in context. They adopt a syntactic perspective.