Sequence Embedding for Zero or Low Resource Knowledge Graph Completion

Knowledge graph completion (KGC) has been proposed to improve KGs by filling in missing links. Previous KGC approaches require a large number of training instances (entity and relation) and hold a closed-world assumption. The real case is that very few instances are available and KG evolve quickly with new entities and relations being added by the minute. The newly added cases are zero resource in training. In this work, we propose a Sequence Embedding with Adversarial learning approach (SEwA) for zero or low resource KGC. It transform the KGC into a sequence prediction problem by making full use of inherently link structure of knowledge graph and resource-easy-to-transfer feature of adversarial contextual embedding. Specifically, the triples ( <h, r, t>) and higher-order triples ( <h, p, t>) containing the paths (p = r(1) -> ... -> r(n) ) are represented as word sequences and are encoded by pre-training model with multi head self-attention. The path is obtained by a non-parametric learning based on the one-class classification of the relation trees. The zero and low resources issues are further optimizes by adversarial learning. At last, our SEwA is evaluated by low resource datasets and open world datasets.