Unsupervised Common Sense Relation Extraction

Abstract

Vast and diverse knowledge about the relations in the world help humans comprehend and argue about their environment. Equipping machines with this knowledge is challenging yet essential for general reasoning capabilities. Here, we propose to apply unsupervised relation extraction (URE), aiming to induce general relations between concepts from natural language. Previous work in URE has predominantly focused on relations between named entities in the encyclopedic domain. The more general, and more challenging, domain of common sense relation learning has not yet been addressed, partially due to a lack of datasets. We present a framework for common sense relation extraction from free-text, associated with two benchmark datasets. We present initial experiments using three state-of-the-art models developed for encyclopedic relation induction. Our results verify the utility of our benchmarks for common sense relation extraction, and suggest ample scope for future work on this important, yet challenging, task.

Tasks

Reproductions