In this experiment, we test the effectiveness of the measure for Causal Association. All the probabilities used in our method are estimated by maximum likelihood estimation from event frequencies in Causal Corpus. To address the data-sparse problem, when we calculate the frequency of an event, we consider those with synonymous arguments refer to the same event. For example, “[Subj teacher][Pre criticize][Obj student]” and “[Subj master][ Pre blame][Obj student]” are
485
synonymous events. Considering the evaluation cost, here we selected 20 groups of synonymous verbs (144 verbs in all) which act as the predicate in the cause event. And we get 6,307 event pairs.
To see the effect of our measure, we compare S with another score PMI. And we calculate these scores respectively using the whole argument structure (-E) or the head verb (-V) as an event.