Abstract: Abstract: Convolutional Neural Network (CNN) based relation extraction has the capability to automatically acquire semantic information from sentences. However, it suffers from the lack of obtaining semantic features at different scales and fails to focus on crucial information. To address these issues, a method for relation extraction based on a multi-scale hybrid attention CNN was proposed. This method leverages multi-scale feature extraction and fusion to obtain finer-grained multi-scale spatial information, enriching the text representation. The hybrid convolution attention mechanism allows the model to adaptively focus on important contextual information.Firstly, The relation extraction is modeled as label prediction in a two-dimensional representation. Next, The combination of attention and convolution is employed to adaptively refine the feature maps. Finally, Two predictors are utilized to jointly predict the relation labels between entity pairs. Experimental results demonstrate that the multi-scale hybrid attention model can capture multi-scale semantic features. Channel attention and spatial attention capture crucial information in channels and spatial locations by assigning appropriate weights, thereby improving the performance of relation extraction. Our model achieves F1 scores of 90.32% on the SemEval-2010 task 8 dataset ,70.74% on the TACRED dataset ,85.71% on the Re-TACRED dataset and 89.66% on the sciERC dataset.

Key words: relation extraction, two-dimensional representation, channel attention, spatial attention, multiscale semantic features

摘要： 摘 要: 基于卷积神经网络的关系抽取能自动获取句子语义信息。针对其缺少不同尺度语义特征信息的获取以及缺少对关键信息关注的问题。提出了基于多尺度混合注意力卷积神经网络的关系抽取方法。该方法通过多尺度的特征信息提取与融合，获得了更细粒度的多尺度空间信息，使得文本表征能力更丰富。混合卷积注意力能关注更关键的信息，使得模型自适应地关注重要上下文信息。首先，将关系抽取建模为二维化表示的标签预测。其次，通过注意力与卷积的结合自适应地细化特征图。最后，使用两个预测器共同预测实体对之间的关系标签。实验结果表明多尺度混和卷积注意力模型能够获取多尺度语义特征信息，通道注意力和空间注意力通过权重来捕捉通道和空间的关键信息，以此来提升关系抽取的性能。本模型在数据集SemEval-2010 task 8、TACRED、Re-TACRED和sciERC的性能F1值分别达到90.32%、70.74%、85.71%和89.66%。

关键词: 关系抽取, 二维化表示, 通道注意力, 空间注意力, 多尺度语义特征