Abstract: Relation classification plays an important part in the structuralization of text data via information extraction. As convolution neural network (CNN) based method can extract high-level features through convolution filters, it achieves competitive performance with other complex-structured networks by only using a standard convolution layer, a pooling layer and a softmax layer. However, when handling samples with long distance between entities, CNN fails to extract effective features, and even extracts the wrong ones from clauses, which results in downgrades of accuracy. Besides, it is observed that existing methods produce inconsistent results when fed with forward and backward instances of the same sample. Therefore, this paper presents a revised CNN based sentence encoder using the shortest dependency paths with selective attention, which only extracts valid sentence features. This paper also introduces the forward instance and backward instance, and incorporates forward and backward instances for relation classification. Experiments verify that the revised relation classification framework and method provide state-of-the-art performance, even without additional artificial features.

Key words: relation classification, selective attention, convolution neural network

摘要： 关系分类是通过信息抽取实现文本数据结构化的重要一环。基于卷积神经网络（convolution neural network，CNN）的关系分类方法，在本身仅包含一个卷积层、池化层和softmax层的情况下，就能取得和其他复杂结构网络相当的效果。但在处理大间距实体的样本时，CNN难以提取有效特征甚至提取出从句中的错误特征，导致分类精度下降。此外，现有方法在输入同一样本的正向实例和反向实例时，会出现结果不一致的情况。针对这两个问题，提出了一种利用最短依赖路径的CNN句子编码器，对句子中与实体联系密切的词语进行选择性注意，增强了CNN抽取特征的有效性；定义了正向实例和反向实例，并设计了一种结合正向实例和反向实例的关系分类框架。实验证明，这种改进的关系分类框架和方法即使没有添加额外特征也取得了领域最优的效果。

关键词: 关系分类, 选择性注意力, 卷积神经网络