作为自然语言处理领域的一项关键任务,事实验证要求能够从大量的纯文本中根据给定的声明检索相关的证据,并使用这些证据推理验证声明。以往的研究通常利用证据句子拼接或图结构表示证据之间的关系,而不能清晰地表示各证据之间的内在关...作为自然语言处理领域的一项关键任务,事实验证要求能够从大量的纯文本中根据给定的声明检索相关的证据,并使用这些证据推理验证声明。以往的研究通常利用证据句子拼接或图结构表示证据之间的关系,而不能清晰地表示各证据之间的内在关联。因此,设计一种基于图谱和文本融合的协同推理网络模型CNGT(Co-attention Network with Graph and Text fusion),以通过构建证据知识图谱和证据句子进行语义融合。首先,根据证据句子构建证据知识图谱,并利用图变换编码器学习图谱表示;其次,利用BERT(Bidirectional Encoder Representations from Transformers)模型对声明和证据编码;最后,通过双层协同推理网络有效地融合推理图谱信息和文本特征。实验结果表明,相较于先进模型KGAT(KnowledgeGraphAttentionneTwork),所提模型在FEVER(FactExtractionand VERification)数据集上的标签准确率(LA)提高了0.84个百分点,FEVER得分提高了1.51个百分点。可见,所提模型更关注证据句子之间的关系,并且通过证据图谱展示出模型对证据句子关系的可解释性。展开更多
In this paper, we first introduce the high efficient pattern matching algorithm named Rete. Second, we analyze the problem appeared in using Rete algorithm to develop expert system. But it develops a new way to realiz...In this paper, we first introduce the high efficient pattern matching algorithm named Rete. Second, we analyze the problem appeared in using Rete algorithm to develop expert system. But it develops a new way to realize Agents' real-time and cooperative reference. Then we put forward to embed the Rete algorithm into Agents and realize cooperative inference by Agents' communication. On these bases, we discuss the accuracy and real-time feature of cooperative inference in detail. Finally, we describe two Agent-communication performatives and a developing instance of multi- agent system to monitor oil tanks.展开更多
文摘作为自然语言处理领域的一项关键任务,事实验证要求能够从大量的纯文本中根据给定的声明检索相关的证据,并使用这些证据推理验证声明。以往的研究通常利用证据句子拼接或图结构表示证据之间的关系,而不能清晰地表示各证据之间的内在关联。因此,设计一种基于图谱和文本融合的协同推理网络模型CNGT(Co-attention Network with Graph and Text fusion),以通过构建证据知识图谱和证据句子进行语义融合。首先,根据证据句子构建证据知识图谱,并利用图变换编码器学习图谱表示;其次,利用BERT(Bidirectional Encoder Representations from Transformers)模型对声明和证据编码;最后,通过双层协同推理网络有效地融合推理图谱信息和文本特征。实验结果表明,相较于先进模型KGAT(KnowledgeGraphAttentionneTwork),所提模型在FEVER(FactExtractionand VERification)数据集上的标签准确率(LA)提高了0.84个百分点,FEVER得分提高了1.51个百分点。可见,所提模型更关注证据句子之间的关系,并且通过证据图谱展示出模型对证据句子关系的可解释性。
文摘In this paper, we first introduce the high efficient pattern matching algorithm named Rete. Second, we analyze the problem appeared in using Rete algorithm to develop expert system. But it develops a new way to realize Agents' real-time and cooperative reference. Then we put forward to embed the Rete algorithm into Agents and realize cooperative inference by Agents' communication. On these bases, we discuss the accuracy and real-time feature of cooperative inference in detail. Finally, we describe two Agent-communication performatives and a developing instance of multi- agent system to monitor oil tanks.
