CaEXR: A Joint Extraction Framework for Causal Relationships Based on Word-Pair Network

Chong Liu,Wenxiang Fang,Debo Cheng,Ruifang Zhai,Li Qin

DOI: https://doi.org/10.1007/978-981-97-5672-8_38

2024-01-01

Abstract:Extracting causal relationships from natural language texts is a fundamental work in constructing eventic graphs. Most previous existing studies are based on the linear representation models of syntactic dependency trees. However, these studies perform poorly for causal relationships extraction in complex contexts, due to the diffcultity of constructing contextual causal information interaction channels, which are crucial for models to understand causal relationships. In addition, existing syntactic dependency tree based models ignore the role of dependency connectivity of causal connectives in syntactic features, which leads to the limited ability of these models to characterise long-distance explicit causal sentences. To address these two issues, we propose a novel joint model named CaEXR. We first reorganize syntactic dependency trees into causal dependency trees according to the causal connectives, which strengthens the long-distance explicit causal features. Then, we transform linear sequence features into high-dimensional word-pair grid features based on a learnable conditional layer normalization module to enchance contextual causal interaction information. Experimental results show that our method achieves the best results on different levels of causal relationships extraction tasks.