Hierarchical Diachronic Embedding of Knowledge Graph Combined with Fragmentary Information Filtering

Knowledge graph(KG) embedding is often used in link prediction, triplet classification, and knowledge graph completion(KGC), which is critical to knowledge relation extraction and recommendation algorithm design. The previous works in KG embedding have made excellent achievements in KGC, but they ignore fragment information and the hierarchy of temporal information in KG. The hierarchy of temporal information indicates that different dimensions of information such as year, month, and day have different degrees of influence on the fact information. Fragmentary information is usually weakly or uncorrelated with the main information, reducing KG embedding learning efficiency. To solve these problems, we propose two methods: Firstly, we propose word frequency filtering to filter fragment information. Secondly, we propose hierarchical diachronic embedding for temporal KG, which provides the characteristics of an entity at different levels of time. We add the consideration of fragmentation filtering and hierarchical temporal embedding to the static KG embedding model SimplE, which results in a novelty model for KGC-HDF-ASimplE(Hierarchical Diachronic Embedding of Knowledge Graph with Fragmentary Information Filtering for Average SimplE). Experimental results manifest that the HDF-ASimplE model is superior to the previous works on KGC, and the evaluation parameters of KGC are increased by 14% on average, Hit@10 to 81.7%. It proves the effectiveness of our method.