A Convolutional Neural Network Method for Relation Classification

Up to now, the relation classification systems focus on using various features generated by parsing modules. However, feature extraction is a time consuming work. Selecting wrong features also lead to classification errors. In this paper, we study the Convolutional Neural Network method for entity relation classification. It uses the embedding vector and the original position information relative to entities of words instead of the features extracted by traditional methods. The N-gram features are extracted by filters in the convolutional layer and the whole sentence features are extracted by the pooling layer. Then the softmax classifier in the fully connected layer is applied for relation classification. Experimental results show that the method of random initialization of the position vector is unreasonable, and the method using the vector and the original position information of words performs better. In addition, filters with multiple window sizes can capture the sentence features and the original location information can replace the complex window sizes.