Generalizable Features Help Semantic Role Labeling
Yang Li · Waseda University Repository (Waseda University) · 2009
In this paper, we take on the challenge of developing effective generalizable features for the task of semantic role labeling in the constituency grammar framework.Based on the knowledge of argument structure, on the constraint imposed by context dependence defined in the theory of argument realization, and on the knowledge of moved and displaced core arguments, we design the base argument configuration (BAC) feature that generalizes across four types of syntactic structures involving moved and displaced core arguments.As part of the effort to derive this base argument configuration feature, we also identify the core and non-core arguments in the system which is the first case in the field of semantic role labeling.Together with two levels of backoff features, the BAC feature effectively solve the argument classification task.However, as the experimental results show, our overall performance is affected by the argument identification module at present.