Relation embedding for open information extraction using a question answering dataset

Abstract

Relation extraction task is a task to extract relation between entities from unstructured text. Due to the lack of labeled data, the majority of relation extraction studies are based on the Distant Supervision. However, it requires predefined relation types and relation instances from Knowledge Base, which causes problems in the scalability of the relation extraction. In order to solve this problem, we propose a PCNN based open relation extraction model trained through Question Answering data set and Open Information extraction with triplet loss. To this end, we hypothesize that queries and corresponding answer sentences share similar relations. Under this hypothesis, the proposed model learns the ability to extract arbitrary relations from (question,answer sentence) pairs of QA dataset. In the experiments, we showed that our hypothesis is valid, and that the proposed model has the ability to extract arbitrary relations without Distant Supervision dataset.