One source of information that can help cross corpus/domain SRL is lexical information. Generalizing lexical attributes extends prior knowledge to a new domain. We have explored two types of lexical generalization. First, co-occurrence data analysis enables the usage of unlabeled data from different sources. The coverage of this method outweighs its inaccuracy. Second, an ontology allows us to provide more guidance for generalization. Again we exploit co-occurrence information to decide the level of reconciliation on the internal structures of the ontology.
We extend the mainline labeling features with generalized lexical information. The effectiveness of each type of extensions is demonstrated by carefully designed experiments and detailed data analysis.