Causal Pattern Representation Learning for Extracting Causality from Literature

Extracting causality from literature has become an important task due to the essential role of causality. Traditional methods use pattern matching to extract causality, requiring domain knowledge and extensive human effort. Recent researches focus on utilizing pre-trained language models due to their success in Natural Language Processing (NLP). However, long sentences in literature hinders the performance of causality extraction. In this paper, we propose to focus on the representation of causal virtual pattern <head_entity, causal_virtual_trigger, tail_entity> and design a Causal Pattern Representation Learning (CPRL) method to tackle this challenge. For the causal_virtual_trigger representation, CPRL applies the attention mechanism on the shortest dependency path between entities to filter irrelevant information. For the head_entity and tail_entity representation, CPRL applies graph convolution networks to encode word dependency on entities. By crawling health-related literature abstracts, we create a new causality extraction dataset, namely HealthCE, with a size of 3479. Experiments on HealthCE demonstrate the effectiveness of our approach over existing causality extraction and general relation extraction baselines on the task of causality extraction.