Causal Relationship Extraction Combined Boundary Detection and Information Interaction

Causal relationship extraction (CRE) is a vital relationship extraction task in natural language processing (NLP), which aims to extract the causal relationship between entities. Existing CRE models have achieved promising performance but suffered with some specific drawbacks. Sequence labeling-based models fail to recognize long entities merely by using the word-level and character-level information, and segment-based models can not capture contextual dependencies and overly focuses on segments rather than specific causal entities. In this paper, we propose a novel causal relationship extraction model CATB (Co-Interactive Adaptive Transformer+Boundary Detection), and tailor for the medical text domain, which independently models the labels and characters of causal events and integrates information to address the limitations of segment-based models. Additionally, we introduce a Boundary Detection Module (BDM) to improve the extraction of long entities in the medical text domain. Experimental results demonstrate that our proposed model significantly outperforms existing methods.