Unsupervised Relation Extraction with Sentence level Distributional Semantics

Abstract

Relation Extraction (RE) aims to identify the relationship between pairs of named entities in natural-language sentences. An unsupervised RE approach extracts relations in the absence of training data. Recently, many state-of-the-art unsupervised approaches have used word embeddings for RE. Such approaches ignore the semantic structure of the complete sentence. On the other hand, in this paper, we propose a novel approach that utilizes sentence encoding for unsupervised relation extraction. Our model classifies the sentence encoding of contextually similar natural-language sentences into clusters using an unsupervised approach, where each cluster consists of one or more potential relations. We queried the cluster for a candidate relation, and used a confidence value/threshold to extract accurate relations without semantic drift. We validated our approach by comparing it with both the unsupervised and bootstrapping approaches. Our experimental results suggest that our model achieves a better F-score on state-of-the-art datasets than the other unsupervised approaches.