Dual Machine Reading Comprehension for Event Extraction

Abstract

Event extraction aims to extract structured triggers and arguments from unstructured text. However, the accuracy of named entity recognition will directly affect the performance of event argument role recognition, which results in error propagation. Meanwhile, treating the event extraction task as a classification task ignores semantic information in sentences. In the paper, we propose a dual machine reading comprehension model (Dual-MRC) for event extraction, which converts the classification task into a span extraction task. The model consists of the part of speech of the candidate argument on the left and the imperative sentence on the right to form a question template, dramatically improving the ability of event extraction. Dual-MRC achieves an F1 value of 74.6% in the event trigger extraction and classification task.Our model performs excellently in the case of data-low scenarios, demonstrating the advantages of machine reading comprehension. Experimental results show that our method is effective on the ACE 2005 dataset, especially for multi-word trigger extraction. In addition, we publish a Chinese mine accident annotation dataset. To the best of our knowledge, this is the first Chinese mine accident event dataset, and we verify the performance of the model in Chinese event extraction on this dataset.