Knowledge element analogy relation recognition using text and graph structure

Knowledge element analogy relation is a corresponding relationship in content, function or other aspects between two knowledge elements. This paper proposes a framework of relation Gaussian processes-based learning for knowledge element analogy relation recognition, which can integrate information from text and relation graph structure. Based on terms or core terms co-occurrence and type compatibility, two rules are first developed to construct candidate analogy relation instances from knowledge element set. Next, three kernels are devised to capture information of terms, semantic types and relative positions of two knowledge elements, and graph Laplacian and expectation propagation algorithm are employed to approximate the relation graph structure. Then, these two types of information are integrated to predict analogy relation. Experimental evaluation on four data sets related to "computer" discipline demonstrates that the rules are effective and integrating three text kernels with relation graph structure can achieve better performance than only text kernels.