TF-ChineseECE: A Chinese Event Causality Extraction Method Combining Roberta and Bi-FLASH-SRU

In response to the excessive number of filled forms in the existing table-based event relation extraction and the inadequate acquisition of table features. In the paper, TF-ChineseECE, a Chinese event causality extraction method combining Roberta and Bi-FLASH-SRU, is proposed. A table filling strategy is developed to transform the tagged relations in the text into a table with labels; the subject and object event features are obtained with the help of Roberta and Bidirectional Built-in Flash Attention Simple Recurrent Unit (Bi-FLASH-SRU), and then the The table feature recurrent learning module is then used to mine the global features, and finally the table decoding is performed to obtain the event causality triad. The experiments are validated using two public datasets in the financial domain, and the results show that the F1 scores of the method proposed in this paper reach 59.2% and 62.5%, respectively, and the Bi-FLASH-SRU model is faster to train and the number of filled forms is less, which proves the effectiveness of the method.