Semantic relation extraction between entity pairs is a crucial task in information extraction from text. In this paper, we propose a new pre-trained network architecture for this task, and it is called the XM-CNN. The XM-CNN utilizes word embedding and position embedding information. It is designed to reinforce the contextual output from the MT-DNN${}_{KD}$ pre-trained model. Our model effectively utilized an entity-aware attention mechanisms to detected the features and also adopts and applies more relation-specific pooling attention mechanisms applied to it. The experimental results show that the XM-CNN achieves state-of-the-art results on the SemEval-2010 task 8, and a thorough evaluation of the method is conducted.

实体对之间的语义关系提取是从文本中提取信息的关键任务。在本文中，我们针对此任务提出了一种新的预训练网络架构，称为XM-CNN。XM-CNN利用单词嵌入和位置嵌入信息。它旨在增强MT-DNN的上下文输出${}_{ķd}$预训练模型。我们的模型有效地利用了实体感知的注意力机制来检测特征，并且采用并应用了更多针对关系的池化注意力机制。实验结果表明，XM-CNN在SemEval-2010任务8上取得了最先进的结果，并对该方法进行了全面评估。