Convolutional Neural Network with SDP-Based Attention for Relation Classification

Abstract

Relation classification plays an important role in the field of natural language processing (NLP). The state-of-theart methods for this task use prior knowledge as features such as WordNet, Part-of-Speech(POS), shortest dependency path (SDP), which is helpful but brings error propagation. In this paper, we propose a convolutional neural network architecture, which builds word-level attention mechanism based on SDP to capture task-oriented patterns in sentences. We explore the way of combining prior knowledge and deep models properly to ease errors in prior knowledge. Additionally, a new objective function is designed to reduce the impact of artificial class, which is seldom touched in previous works. Experiments on the SemEval-2010 Task 8 benchmark dataset show that our model outperforms some of the state-of-the-art methods.