Embedding the Semantic Knowledge in Convolution Kernels

Convolution kernels, such as tree kernel and subsequence kernel are useful for natural language processing tasks. However, most of them ignore the semantic knowledge. In order to solve the problem, this paper proposes a new method to embed the semantic knowledge into kernel calculation. The new method has been applied to extract the ORG-affiliation relation from Chinese texts and achieves an average F-measure of 82.1%. Comparing with feature-based method and the traditional Word-sequence kernel, it provides significant improvement.