Distant Supervised Relation Extraction Based On Recurrent Convolutional Piecewise Neural Network

Abstract

Distant supervised relation extraction (RE) is currently an effective way to solve the problem of extracting relation from large amounts of unlabeled data.The purpose of distant supervised relation extraction is to identify the relation between the two entities marked in a sentence. However, there are two existing problems.The one is that some methods need to draw on entity description of external knowledge base to enrich entity information, but in reality not every time we have entity description of an external knowledge base. The other one is that the effects of distant supervised relation extraction are not very ideal currently.This paper proposes a novel relation extraction model based on recurrent piecewise convolutional neural network structure to solve the problems above. Firstly, based on the recurrent convolutional neural network structure, the embedding of every word in a sentence is added with context information to enrich the characteristics of the words. And then with piecewise max pooling, it captures the information throughout the entire sentence. Secondly, the semantic information of a sentence can indirectly reflect the relation of the entity.This paper employs sentence vectors to add the semantic information to improve the accuracy of the distant supervised relation extraction. The experimental results are based on real-world dataset. Our model makes full use of the information characteristics of the dataset and has great improvement on the real-world dataset.It proves that our model in this paper exceeds various baselines.