Unsupervised Relation Extraction Using Sentence Encoding

Relation extraction between two named entities from unstructured text is an important natural language processing task. In the absence of labelled data, semi-supervised and unsupervised approaches are used to extract relations. We present a novel approach that uses sentence encoding for unsupervised relation extraction. We use a pre-trained, SBERT based model for sentence encoding. Our approach classifies identical sentences using a clustering algorithm. These sentences are used to extract relations between two named entities in a given text. The system calculates a confidence value above a certain threshold to avoid semantic drift. The experimental results show that without any explicit feature selection and independent of the size of the corpus, our proposed approach achieves a better F-score than state-of-the-art unsupervised models.