Combining Position-aware CNN and RNN for Relation Extraction

Abstract

Recurrent neural network (RNN) is often used for relation extraction that is based on long range dependency analyses and connections between nodes within the given sentences. However, RNN has a weaker ability in extracting certain important n-grams from the given datasets compared with convolutional neural network (CNN). Therefore, the combination of CNN and RNN becomes necessary for higher quality of relation classification. We propose Position-aware CNN (PA-CNN) which incorporates position information into CNN through an attention mechanism. The proposed PA-CNN is able to capture important linguistic clues for relation extraction effectively. Furthermore, we combine Position-aware CNN and RNN to take the benefits of both CNN and RNN to improve the relation extraction performance. Experiments on TACRED, a complex relation extraction dataset, show that our proposed Position-aware CNN outperforms traditional CNN, and the combinations of CNN and RNN outperform single CNN or RNN models for relation extraction.