双关系预测与特征融合的实体关系抽取模型

Entity relation extraction model with dual relation prediction and feature fusion

现有分阶段解码的实体关系抽取模型仍存在着阶段间特征融合不充分的问题,会增大曝光偏差对抽取性能的影响。为此,提出一种双关系预测和特征融合的实体关系抽取模型(entity relation extraction model with dual relation prediction and feature fusion,DRPFF),该模型使用预训练的基于Transformer的双向编码表示模型(bidirectional encoder representation from transformers,BERT)对文本进行编码,并设计两阶段的双关系预测结构来减少抽取过程中错误三元组的生成。在阶段间通过门控线性单元(gated linear unit,GLU)和条件层规范化(conditional layer normalization,CLN)组合的结构来更好地融合实体之间的特征。在NYT和WebNLG这2个公开数据集上的试验结果表明,该模型相较于基线方法取得了更好的效果。

The staged decoding entity relation extraction model still has an insufficient feature fusion problem between stages,which increases the impact of exposure bias on the extraction performance.Herein,we propose a new entity relation extraction model with dual relation prediction and feature fusion(DRPFF).DRPFF uses a pretrained model of bidirectional encoder representation from transformers to encode texts,and a two-stage dual relation prediction structure is developed to reduce the false triples’generation.Between stages,a structure combining gated linear units and conditional layer normalization is utilized to fuse features better between entities.Experimental findings on two public datasets,NYT and WebNLG,demonstrate that the presented method has better results than the baseline methods.