基于对抗学习和全局知识信息的关系检测技术研究

针对现有的知识库关系检测任务对于一些不可见关系无法做到准确的向量表示而出现词汇溢出的问题,提出了基于对抗学习和全局知识信息的关系检测模型。该模型使用对抗学习对知识库关系表示模型进行特征强化,使用TransH(translating on hyperplanes)模型提取全局知识信息,同时通过联合训练,将全局知识信息融合进关系表示模型中,进一步提升关系模型的表示能力。实验结果表明,提出的融合模型对于关系检测效果有一定的提升,并且缓解了词汇溢出的问题。

Research on relation detection based on adversarial learning and global knowledge information

Aiming at the problem of out-of-vocabulary due to the fact that the existing knowledge base relationship detection task are unable to achieve accurate vector representation for some invisible relationships,this paper proposed a relationship detection model based on adversarial learning and global knowledge information.The model used adversarial learning to enhance the characteristics of the relational representation model of the knowledge base,and used the TransH model to extract global knowledge information.At the same time,through joint training,it integrated the global knowledge information into the relatio-nal representation model to further improve the representation ability of the relational model.The experimental results show that the proposed fusion model improves the effect of relationship detection and alleviates the problem of out-of-vocabulary.