基于注意力机制与PCNN的地质关系抽取方法

Geological Relation Extraction Method Based on Attention Mechanism and PCNN

地质领域的关系抽取对地质数据的智能分析以及地质领域知识图谱的构建具有非常重要的意义。针对地质领域关系抽取任务中数据集缺失,特征提取困难等问题,论文基于远程监督思想构建大规模地质关系抽取数据集,提出注意力机制与分段卷积神经网络(PCNN)相结合的关系抽取模型。该模型使用分段卷积神经网络自动提取训练实例的语义特征,在分段卷积神经网络中加入分段注意力机制突出实例中重要的分段,增强模型对实例中重要特征的提取能力,并引入句子级注意力机制降低由远程监督带来的错误标注数据的影响。在基于远程监督思想构建的地质领域关系抽取数据集上的实验结果表明,论文模型的准确率,召回率和F1值均高于其他的基线模型,具有更好的关系抽取能力。

Relation extraction in the geological domain is of great significance to the intelligent analysis of geological data and the construction of knowledge graph of Geology.To solve the problems of missing data sets and the difficulty in feature extraction for relation extraction tasks of the geological domain,this paper builds a large-scale relation extraction data set in geological domain based on the idea of distant supervision,and proposes a distant supervision relation extraction model which combines attention mechanism and piecewise convolutional neural network(PCNN).The model in this paper uses the piecewise convolutional neural network to automatically extract the semantic features of the training instances.At the same time,the segmented attention mecha-nism is added to the piecewise convolutional neural network to highlight important segments in the instance,which strengthens the model's ability to extract important features in training instances.In addition,a sentence-level attention mechanism is introduced in-to this model to reduce the impact of incorrectly labeled data due to the distant supervision.The experimental results on the relation extraction data set in the geological field based on distant supervision show that the precision,recall,and F1 value of this model are all higher than other baseline models,which illustrates that the method has better relation extraction capabilities.