Tree Kernel-based Relation Extraction with Various Entity-Related Features

This paper proposes a convolution tree kernel-based approach for relation extraction where parse trees are expanded with various entity-related features, such as entity type, subtype, and mention level. Our study indicates that not only can our method effectively capture both syntactic structure and entity information of relation instances in a single tree kernel, but also can avoid the difficulty with tuning the parameters in composite kernels. Moreover, we apply several linguistic rules to further prune out noisy information from the parse trees representing relation instances. Empirical evaluation on the ACE 2004 benchmark corpus shows that our system works well and achieves promising performance compared with other kernel-based systems.