A Re-examination of Dependency Path Kernels for Relation Extraction

Extracting semantic relations between entities from natural language text is an important step towards automatic knowledge extraction from large text collections and the Web. The state-of-the-art approach to relation extraction employs Support Vector Machines (SVM) and kernel methods for classification. Despite the diversity of kernels and the near exhaustive trial-and-error on kernel combination, there lacks a clear understanding of how these kernels relate to each other and why some are superior than others. In this paper, we provide an analysis of the relative strength and weakness of several kernels through systematic experimentation. We show that relation extraction can benefit from increasing the feature space through convolution kernel and introducing bias towards more syntactically meaningful feature space. Based on our analysis, we propose a new convolution dependency path kernel that combines the above two benefits. Our experimental results on the standard ACE 2003 datasets demonstrate that our new kernel gives consistent and significantly better performance than baseline methods, obtaining very competitive results to the state-ofthe-art performance.