Dual decomposition method for chinese predicate-argument structure analysis

Yanyan Luo, Masayuki Asahara, Yuji Matsumoto

Abstract: Most of the previous work on predicate-argument(PA) structure analysis utilized pipeline strategies with several subtasks. However, when the subtasks are sequentially processed, the errors propagate from the preceding subtasks to consecutive subtasks. In this paper, the PA structure analysis is divided into three subtasks: predicate sense disambiguation, argument classification and argument identification. A dual decomposition method is used to join the latter two subtasks, which alleviates the error propagation by making the decisions from the two subtasks as consistent as possible. Furthermore the predicate sense disambiguation task is integrated into the argument classification task by handling syntactic dependencies and semantic dependencies between a predicate and its arguments. Experiments show that our approach achieves competitive results compared to the state-of-the-art systems without special feature selection procedures.