An attention-based model for joint extraction of entities and relations with implicit entity features

Abstract

Extracting entities and relations is critical to the understanding of massive text corpora. Recently, neural joint models have shown promising results for this task. However, the entity features are not effectively used in these joint models. In this paper, we propose an approach to utilize the implicit entity features in the joint model and show these features can facilitate the joint extraction task. Particularly, we use the hidden-layer vectors extracted from a pretrained named entity recognition model as the entity features. Thus, our method does not need to design the entity features by hand and can benefit from the new development of named entity recognition task. In addition, we introduce an attention mechanism in our model which can select the informative parts of the input sentence to the prediction. We conduct a series of experiments on a public dataset and the results show the effectiveness of our model.