摘要
事实核查是指基于证据文本的虚假信息检测任务,目前已有的研究方法主要是将声明文本与证据文本拼接后输入预训练模型进行分类判断,或者通过单一节点的全连接图进行推理判断.这些方法忽略了证据文本间的远距离语义关联和其包含的噪声干扰.针对以上问题,该文提出了一种基于跨证据文本实体关系的图卷积神经网络模型(Cross-Evidence Entity Relation Reasoning Model,CERM).该模型以多个证据文本的实体共现关系为基础,聚合不同实体对象的语义结构信息,同时减小噪声信息干扰,有效提升模型的虚假信息判别能力.实验结果证明,在公开数据集上该文提出的方法在通用评测指标上均优于现有的对比模型,验证了 CERM模型在事实核查研究任务上的有效性.
Abstract
Fact checking is defined as the task of detecting false information based on evidence.To address the long-range semantic association between evidences,this paper proposes a cross-evidence entity relation reasoning model(CERM for short).The CERM model constructs a graph neural network centered on the entity relationship between evidence,and aggregates the semantic structure information of the same entity by the same entity link between dif-ferent evidence texts.Experiments on a public fact verification benchmark show that the proposed model is superior to the existing models in general evaluation indicators.
基金项目
国家重点研究与发展计划(2020YFB1406902)