Designing a proper scoring function is the key to ensuring the excellent performance of knowledge base (KB) embedding. Recently, the scoring function search method introduces the automated machine learning (AutoML) technique to design the task-aware scoring function for any given binary relational data (a.k.a. knowledge graph, KG), which achieves state-of-the-art performance. However, the efficiency and effectiveness of the current searching method are still not as good as desired. First, the existing method consumes a lot of computational overhead to search for a proper scoring function. Second, the existing method can only search the scoring function for the given binary relational data, which is a special form of general KBs (i.e., N-ary relational data).

In this thesis, we present...[ Read more ]

Designing a proper scoring function is the key to ensuring the excellent performance of knowledge base (KB) embedding. Recently, the scoring function search method introduces the automated machine learning (AutoML) technique to design the task-aware scoring function for any given binary relational data (a.k.a. knowledge graph, KG), which achieves state-of-the-art performance. However, the efficiency and effectiveness of the current searching method are still not as good as desired. First, the existing method consumes a lot of computational overhead to search for a proper scoring function. Second, the existing method can only search the scoring function for the given binary relational data, which is a special form of general KBs (i.e., N-ary relational data).

In this thesis, we present three steps to progressively perform the automated scoring function design for knowledge base embedding. First, we propose ERAS, an efficient scoring function search method on binary relational data. We suggest sharing the em- beddings among candidate scoring functions to avoid repeated embedding training in literature, which accelerates the search procedure. However, ERAS cannot well adapt to the more complex case, N-ary Relational Data. Therefore, we next propose S2S to extend the scoring function search from the binary to the N-ary scenario. S2S upgrades the search space of scoring functions and improves the search algorithm. Finally, we rethink the data sparsity issue in the KB embedding. Compared with other modelings, representing KBs with multi-relational hypergraphs is a more natural way to encode facts with different arities. Therefore, we propose a unified framework to automatically design the suitable graph neural network, especially the message function, for any given KBs.