基于预训练模型的军事领域问答关系检测方法

为解决军事领域问答服务中用户问询语句的关系链接错误问题,提高知识库问答的准确性,提出了一种基于预训练语言模型的关系检测方法。首先,摒弃了用户问句中的实体名称信息,加入约束性本体信息,结合预训练语言模型嵌入注意力机制进行了关系检测模型研究;然后,将该关系检测方法结合军事语料应用于军事知识库问答任务中,进行了实验验证。结果表明:约束性本体信息的加入扩展了本体层级信息量与本体知识拓扑结构,对关系检测结果进行了约束,测试关系链接精准率提升了6.2%左右;预训练模型的嵌入为军事数据注入了更多背景知识,相比于未嵌入前,训练精准率提升了10%左右,说明结合知识库的信息特征,整体增强了关系检测在军事领域问答服务中的实际应用效果。

基于概率软逻辑模型的实体解析

在马尔科夫逻辑网(MLN)的实体解析算法中任意闭原子采用硬约束,导致推理及权重学习过程较难收敛到最优解,降低解析精度及执行效率。为此,提出一种将概率软逻辑(PSL)模型应用到实体解析中的方法,该模型中闭原子采用软约束,易于进行知识推理与权重学习。阐述PSL模型基本理论,通过实体关系、实体属性、本体约束构造PSL模型的逻辑规则,描述实体解析的匹配过程,根据PSL模型的推理机制实现实体解析的决策过程。实验结果表明,与基于MLN的实体解析算法相比,该方法可大幅提高实体解析的准确率、F1值及执行效率。

Combining ODM and OCL in ontology verification

In order to automatically find and conveniently rectify the structural conflicts appearing in a new ontology model version after a series of ontology evolution,we propose an ontology model verification approach,in which the object constraint language (OCL) and an ontology definition meta-model (ODM) are used to complete the ontology verification.The ODM is composed of the ontology related elements and the definition rule related elements.The OCL is employed to describe the elements in the ontology definition meta-model:OCL extension for the ontology related elements,and OCL extension for the ontological design patterns.All of the above OCL rules will describe the constraint relationships between ontology elements.Associated with an example,the application of the ontology model verification approach based on OCL and ODM is introduced.Consequently,the conflicts happened in the ontology changing and evolution can be easily verified and rectified through this approach.