Event causality extraction based on Transformer sequence annotation model

Abstract

Text data such as research reports and announcements in the financial field contain a large amount of event causality that can be extracted and thus applied to downstream tasks such as prediction and Q&A. Traditional event causality extraction methods extract through sentence templates, which cannot cope with multiple pairs of causality in a sentence. This paper considers the event causality extraction task as a sequential annotation task. The event causality labels are divided into ”core noun in the cause”, ”predicate or state in the cause”, ”central word”, ”core noun in result”, and ”predicate or state in result”. We proposed using the Transformer sequence annotation model based on lexicon matching to identify and extract event causality. The F1 value of the Transformer model reaches 58.70 %, and the F1 of BERT+Transformer comes the highest, 69.49 %.