Relation Extraction - A Study on Kernel Functions

This paper presents a study on one of the important methods for the relation extraction problem. Roughly, the problem of relation extraction is the problem of finding all (some) facts in a given natural language sentence. The method is based on kernel functions for the support vector machines (SVM) algorithm. A kernel function is a function that represents the similarity of semantics between sentences. The kernel functions from the literature considered, in this paper, to be the most important ones are: bag-of-words kernel (BOWK), global context kernel (GCK), and sub-tree kernel (STK). The SVM algorithm is utilized to learn a candidate sentence classification model for a target relation
$r$
. The classification model can then be used to determine whether a new candidate sentence (with two labeled object descriptions) is a positive or a negative candidate sentence with respect to the given relation
$r$
. The paper includes a systematic comparison of the various kernel functions proposed in the literature. Based on the strength and weaknesses of the various kernel functions, new kernel functions are proposed, namely, directed global context kernel (DGCK), global context synonyms kernel (GCSK), directed global context synonyms kernel (DGCSK), and sub-tree synonyms kernel (STSK). Lastly, the paper contains a comparison of the evaluation results of the new kernel functions with the results of the kernel functions proposed in the literature.