Good Negative Sampling for Triple Classification

Knowledge graphs are large and useful sources widely used for natural question answering, Web search and data analytics. They describe facts about a certain domain of interest by representing them using entities interconnected via relations in the way of triples. However, due to the fact that they are created under the Open World Assumption, they are incomplete. Knowledge graph completion includes the triple classification task, for discerning correct from incorrect triples. In this regard, knowledge graph embedding models have been proposed for the knowledge graph completion tasks. However, knowledge graphs include only positive triples and training models with only positive triples over generalize, therefore, these models require negative examples. A random negative sampling generates low-quality negative triples which give rise to the zero loss problem during training. In this work, Good Negative Sampling, which is a negative sampling strategy that aims to improve the negative generation process by using background ontological knowledge is put forward. We prove our strategy on a state-of-the-art embedding model - KG-BERT for the triple classification task - on a benchmark dataset - FB13. As result, we demonstrate that the Good Negative Sampling strategy overcomes other state-of-the-art negative strategies, with significant differences.