神经网络知识图谱模型在机器翻译知识图谱推理中的应用研究

Research on the Application of Neural Network Knowledge graph Model in Machine translation Knowledge graph Reasoning

知识图谱通过结构化方式完成实体与实体信息之间关系的存储,能够快速高效地处理信息,但存在稀疏性、不完整性等缺点。研究首先通过设计可逆残差模块、输入模块和输出模块构建高效率神经网络(Highly Efficient Transformer,HET),随后利用FOL规则准确率高且可解释性强的特点完成知识推理,由此形成混合逻辑规则与神经网络的知识推理模型(Highly Efficient Transformer Inductive Learner,HETIL)。结果表明,随着步数的增加,HET高效神经网络模型的困惑度最低仅为3.6。在英语机器翻译实例中,HETIL混合推理模型对网页英语翻译场景的平均倒数排名(Mean Reciprocal Ranking,MRR)指标可达0.67,耗时最低仅为18 s。说明所提出模型在机器翻译中具有较高的运行效率,能够有效完成推理任务,为英语机器翻译领域的发展提供了新的思路和方法。

The Knowledge graph stores the relationship between entities and entity information in a structured way,which can process information quickly and efficiently,but it has shortcomings such as sparsity and incompleteness.The study first constructs a High Efficient Transformer(HET)by designing reversible residual modules,input modules,and output modules.Then,it utilizes the high accuracy and interpretability of FOL rules to complete knowledge inference,forming a mixed logic rule and neural network knowledge inference model(HETIL).The results show that with the increase of the number of steps,the lowest Perplexity of HET efficient neural network model is only 3.6.In the case of English Machine translation,the Mean reciprocal rank(MRR)index of HETIL hybrid reasoning model for web page English translation scenarios can reach 0.67,and the minimum time consumption is only 18 seconds.It shows that the proposed model has high running efficiency in Machine translation and can effectively complete reasoning tasks,which provides new ideas and methods for the development of English Machine translation.