基于树核的中英文语义关系抽取比较

urrently kernel-based approaches have achieved promising performance for Chinese-English semantic relation extraction. Therefore presentation of structured information and parsers to generate them are crucial to kernel-based Chinese-English semantic relation extraction. This paper systematically compares three kinds of structured information (Minimum Complete Tree, Shortest Path-enclosed Tree, and Dynamic Relation Tree) and three kinds of parsers (Stanford, Berkeley, and Charniak) on the performance of the task of Chinese-English relation extraction. Particularly we employ the approximate random test method for significance tests in order to statistically discern the difference between various kinds of structured information and parsers. The experiments on the ACE 2004 English and ACE 2005 Chinese corpora show that Dynamic Relation Tree performs best and both the Charniak and Berkeley parsers outperform the Stanford parser.