We evaluate event-argument extraction approach by comparing with two other automatic methods used in previous works. One is relied on semantic role labeling (SRL) technique to identify the subject (Arg0) and the object (Arg1) referred in (Riaz, 2010). The other one used structure-mapping rules based on whole dependence parser, which is similar to (Khoo, 2000). The difference between the Whole Parser method and Local Parser method is whether parsing the causal expression or respectively parsing the two event expressions.
In order to reduce evaluation costs, we randomly select 5,000 sentences from the Causal Corpus as the test data-set. Using a Chinese NLP Toolkit (2011), we get the whole dependency trees, local dependency structures and semantic role labeling results of these sentences. In our approach, we use 36 predefined rules to extract event-argument structures. We note that, if an event expression contains multiple contiguous verbs, then its dependency tree is usually ambiguous. To improve the precision, we remove some obvious wrong dependency trees using heuristic rules.