Knowledge enabled relation extraction

Abstract

Relation extraction is the task of extracting relationships between entities from text, where the text can be a part of a sentence, a document, or across multiple documents. This task has been popular for decades and is still of keen interest. Various techniques have been proposed to solve the relation extraction problem, among which the most popular are using distant supervision, deep learning-based models and reasoning-based models. However, these techniques rely on the knowledge between entities in a text and do not consider the background knowledge of the entities themselves, such as the entity type, synonyms and entity definitions. Predicting relations based on this knowledge is challenging due to the latent and unspecific contexts that introduce noise. To address these issues, we investigate mechanisms to incorporate background knowledge into the relation extraction task. We consider publicly available and relevant ontologies and knowledge graphs as sources of background knowledge. We propose three approaches named ReOnto, DocRE-CLip, and KXDocRE for relation extraction from text at three levels of granularity (sentence, document and across documents). These approaches embed knowledge in deep learning based models, and this has led to an improvement in their performance. We evaluate our approaches using domain-specific and general datasets. The results validate the utility of considering background knowledge for relation extraction.