Motivated by the needs of generic methods to acquire specific relations between events, we explore an automatic three-phased approach. We take the causal relation as example in this paper. We first use lexico-syntactic patterns to not only recognize causal relations from the web text, but also identify pairs of event expressions. Then, we extract the predicate-argument structure of each event expression based on its dependency parser tree in local scale. At the last step, we propose a statistical score S to measure the causal association between potential related events, and prune relations with low S value. The experimental results have shown that (a) the use of local dependency tree extensively improves both the accuracy and recall of event-arguments extraction task; (b) our measure which is an improvement of PMI has better performance.
There are two interesting directions in the future. First, identifying causality boundary automatically rather than just using separators in a pattern. Second, the three arguments referred in our work are not enough in some special cases. For example, the effects of events “he works carelessly” and “he works carefully” are commonly different. To distinguish from these causes, we will introduce new arguments as the solution.