概率条件下基于双目标交替优化的知识表示模型

A Knowledge Representation Model Based on Bi-Objective Alternate Optimization Under Probability

针对Trans D模型参数多和实体两种表示间没有关联的问题,提出一种改进的知识表示模型PTrans D。通过减少实体投影数,并对实体进行聚类来减少参数量,同时利用K-L(Kullback-Leibler)散度限制实体投影和对应实体类,使其概率分布相同。在模型训练时,对三元组损失和K-L损失交替优化,从类间距大的实体类中替换实体,提高负例质量。最后,在知识图谱数据集上进行三元组分类和链接预测实验。结果表明,该模型的性能在各项指标上均有明显提高,可以应用于知识图谱的完善和推理等。

Aiming at the problem that the TransD model has many parameters and the two representations of entities are not related, an improved knowledge representation model PTransD is proposed, which reduces the number of parameters by reducing the number of entity projections and clustering entities, while using K-L(Kullback-Leibler) The divergence limits the entity projection to the same probability distribution as the corresponding entity class. During model training, the triple loss and K-L loss are alternately optimized, and the entities in the classes with large spacing between the entities are replaced to improve the quality of negative examples. Finally, based on the experimental results of triple classification and link prediction on the knowledge graph data set, the performance has been significantly improved in various indicators. It can be applied to the perfection and reasoning of knowledge map.