Event Causality Extraction Based On Fusion Attention

Relationship extraction is an important task in the field of information extraction in natural language processing. It refers to judging the logical relationship between events, including co-referential, subordinate, temporal and causal relationships. Causality can be divided into explicit causality and implicit causality. At present, it is difficult to extract the implicit causality between events in natural language processing models. In addition, it is often difficult to extract the long-distance causality in causality extraction. To solve these problems, event causality extraction based on fusion attention (ECEFA) is proposed to transform event causality extraction into causal argument extraction, taking into account sentence level context information and document level context information. ECEFA uses Albert to generate word embedding, which can alleviate the common problem of polysemy in Chinese and improve the semantic understanding ability of the model. Dependency syntax is introduced to solve the problem of long-distance dependence in causal argument extraction and avoid the omission of implicit causal events. The experimental results on the CEC2.0 Chinese corpus show that the proposed model improves 2.54%, 2.83% and 2.69% in precision, recall and F1-score, respectively, compared to the best baseline model CSNN. On the dam operation log corpus, the improvement over the best baseline JMCEE was 3.25%, 3.12% and 3.18% in precision, recall and F1-score respectively.