However, the distribution of timestamps in the knowledge graph is unbalanced, and each fact is projected to all the timestamps included in its time span, which results in the “overlapping part timestamps” between different facts for the same completion task. The unbalanced distribution of timestamps will aggravate this problem.įor example, given a series of temporal knowledge graph completion tasks ( C h a r l e s _ D a r w i n, h a s W o n P r i z e, R o y a l _ M e d a l, ), ( A r t h u r _ S m i t h- _ W o o d w a r d, h a s W o n P r i z e, R o y a l _ M e d a l, ), ( H a n s _ A d o l f _ K r e b s, h- a s W o n P r i z e, R o y a l _ M e d a l, ), where the end time of triples is replaced with the default maximum value due to being unknown, there will be multiple triples with the same relations and tail entities to be projected onto the same timestamp hyperplane in the time of “overlapping part”. And the training of the head entity will be indistinguishable. When calculating the score for head prediction, the parameter Dasgupta feeds is the number of the timestamp hyperplane where the start time (such as 1853, 1917, 1954) of the triples is located. But it is obvious that a triple is valid for a range of timestamps. The purpose of this is to avoid the above influence as much as possible, because the timestamp of the start time is less likely to have confused triples. The prediction of the earlier triples will be more accurate, and the prediction of the later triples will be more ambiguous, because the timestamp hyperplane, in which the later triple’s start year is located, will contain more confused triples. Although Tang tries to solve this problem, it does not directly model in the embeddings to distribute the timestamps evenly. Instead, it uses an external unit (design a timespan gate in GRU ) to solve this problem, which greatly increases the complexity of the model.Īlthough these models have good results in the experiments and datasets they show, they all ignore the time information of the knowledge graph. There are few explorations about the completion methods of temporal knowledge graphs. Jiang provides a link prediction strategy by modeling the sequence of temporal relations, and Leblay studies various methods of the interaction between time and relation through learning the embedding of them. Dasgupta proposes a model directly encoding time information in the embedding, which gives us great enlightenment.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |