【摘要】 针对现有小样本关系抽取模型样本嵌入包含的信息量少、参考大量无关的特征，关系抽取效果不佳的问题，提出了EMD＿ProtoNet模型。利用BERT进行样本嵌入，采用原型网络(Prototypical Networks, ProtoNet)为各个关系类别计算类原型，使用土方移动距离(Earth Mover’s Distance, EMD)作为距离度量在匹配代价最小的样本嵌入之间生成最优匹配，通过计算BERT样本嵌入之间的距离确定样本的相关性，根据相关性进行抽取。采用一种交叉参照机制生成EMD公式中节点的重要性权重，从而更多地关注具有较大区别性的特征。实验结果表明，EMD＿ProtoNet能够有效的表示样本嵌入并且有效的度量距离，具有更高的准确率和更快的收敛速度，适用于小样本关系抽取任务。更多还原

【Abstract】 To solve the problem that the existing model of few-shot relation extraction contains less information, refers to a large number of irrelevant features, and has poor effect of relation extraction, an EMD＿ProtoNet model is proposed.Samples are embedded by BERT, ProtoNet is used to calculate the class prototype for each relationship category, and EMD is used as the distance measure to generate the best match between samples embedded with the lowest matching cost.The correlation of samples is determined by calculating the distance between BERT samples embedded, and the samples are extracted according to the correlation.A cross-reference mechanism is used to generate the importance weights of nodes in the EMD formula, so as to pay more attention to the features with great differences.Experimental results show that EMD＿ProtoNet can effectively represent sample embedding and measure distance, with higher accuracy and faster convergence speed, and is suitable for few-shot relation extraction tasks.更多还原