Research On Tree Kernel-based Semantic Relation Extraction Between Named Entities

Semantic relation extraction (SRE) between named entities plays an important role in natural language processing, especially information extraction. It also has broad application prospects. As a hot research topic in the machine learning community, tree kernel methods have been successfully applied in SRE to a certain extent. As an alternative to feature-based methods, tree kernel methods provide an elegant solution to explore implicitly structured features by directly computing the similarity between two trees.This paper proposes a tree kernel method in SRE based on the convolution tree kernel. Evaluation on the ACE2004 corpus shows that our method achieves the F-measures of 72.5 and 57.4 over relation detection and classification respectively.In order to further improve the performance, this paper also proposes an expanding and pruning strategy to keep more semantic information and filter out unnecessary noisy information. Evaluation on the ACE RDC 2004 benchmark corpus shows that our strategy significantly improves the performance and achieves the F-measures of 79.1 and 71.9 in relation detection and classification respectively.Finally, this paper adopts the RST theory and proposes a novel discource analysis tree to deal with complicated and low-performed long-distance relation instances. Here, the discourse analysis tree is further combined with the original parse tree in a novel way. Tentative evaluation on the ACE RDC 2004 corpus shows that the discourse analysis tree structure is useful on extraction of long-distance relations and improves the performance, though the discourse analysis tool SPADE applied in this paper is far from perfect.