Joint Model of Triple Relation Extraction with Label Embeddings

Abstract

Extracting triple relations from unstructured text is an important task for building a knowledge graph. The method for triple relation extraction can be broadly divided into the pipeline and joint extraction models. Both methods focus only on word embedding, and treating labels as meaningless information. These labels are effective intermediate features that capture the semantic rules between words. In this paper, with the joint extraction mode of relation extraction, a framework for joint extraction of triples based on label embedding is proposed, called TriLab. Specifically, it first uses a pretrained model to convert the text into a vector representation. Then, the decoding part implements the extraction of triples through two two-class classifiers, sharing the encoding layer. The label-word joint embedding problem is regarded as the focus of the classification task. The experiments were conducted on a large-scale Chinese data set. The comparative results show that the TriLab framework outperformed the previous methods in triple extraction.