Two-phased event causality acquisition: Coupling the boundary identification and argument identification approaches

Abstract

Event causality is indispensable for knowledge-driven intelligent systems. In this paper, we propose a supervised method of extracting event causalities such as forest is cut down→forest is destroyed from web text. While relation identification using lexico-syntactic patterns (LSPs) is not novel, it is still challenging to extract the event expressions with necessary arguments from identified causality mentions. To address this issue, our method divides event-pair extraction into two phases: event boundary identification and missing argument identification. In the first phase, we propose a Naive Baysian probability method to identify the boundary of causal events, and extract the corresponding text fragments as event expressions. Secondly, we learn a multi-class decision tree (LADTree) to identify the missing argument for each incomplete event. Experimental results showed the good effectiveness of our approach on a large-scale open corpus.