Causality Extraction based on Sequence Tagging with Triple Attention for Cyber Physical Applications

Causality extraction from noisy texts for cyber physical applications has long been a challenging task. In this paper, we propose an enhanced BiLSTM-CRF sequence tagging model that achieves state-of-the-art results. Specifically, we extend the multi-head self-attention mechanism by integrating the Coordinate Embedded Channel Attention (CECA) module into the Scaled Dot-Product Attention. Within the CECA module, we innovatively formulate the forward and backward outputs of the BiLSTM layer as two channels to make flat text three-dimensional, subsequently enhancing attention mechanisms inspired by image processing techniques. Experiments conducted on the extended SemEval 2010 Task 8 dataset have demonstrated the effectiveness of our approach, achieving a precision of 84.46%, a recall of 84.79% and an F1-score of 84.62%, surpassing all other comparative models.