Event causality identification by modeling events and relation embedding

Abstract

Events and event relations contain high-level semantic information behind texts. In this paper, we mainly discuss event causality relation identification. Traditional approaches of causality relation identification rely on the recognition of casual relationship connectives or manual features of causality relationships, and these methods have disadvantage of low recognition coverage and being lack of adaptive. To solve this problem, we propose a novel model based on modeling event and event relation. We use word sequence around event trigger as input data and use event based Siamese Bi-LSTM network to model events by encoding the event representations into a fixed size vectors, and then these events representations are applied in relation embedding training and prediction. Experimental results show that the proposed method can achieve better effect on CEC 2.0 corpus.