Causal Relationship Extraction Using BiGRU-Attention-CRF with Character Features and Iterated Dilated CNNs

Abstract

Causality extraction is an essential task in natural language processing, aiming to identify and structure causal relationships from unstructured text. While deep learning techniques have advanced this field, challenges remain in extracting causal relationships effectively, particularly in representing lexical features and capturing multiple and overlapping causal relations. This paper proposes a novel method for causality extraction using a BiGRU-Attention-CRF model with enhanced feature representation. To improve the extraction performance, character features are integrated with pre-trained embeddings, and Iterated Dilated Convolutional Neural Networks (IDCNNs) are employed to capture local information while minimizing information loss. The BiGRU-Attention mechanism is utilized to extract deeper contextual information and improve training speed compared to traditional BiLSTM models. Experimental results demonstrate that the proposed method achieves competitive performance in causality extraction tasks, offering efficient training and robust feature extraction. However, the model's performance is limited by the dataset size, and it is currently confined to intra-sentence causality extraction. Future work will explore enhancing cross-sentence and document-level causality extraction using graph convolutional networks and low-resource question templates.