ReLink: Open Information Extraction by Linking Phrases and Its Applications

Abstract

Recently, many Open IE systems have been developed based on using deep linguistic features such as dependency-parse features to overcome the limitations presented in older Open IE systems which use only shallow information like part-of-speech or chunking. Even though these newer systems have some clear advantages in their extractions, they also possess several issues which do not exist in old systems. In this paper, we analyze the outputs from several popular Open IE systems to find out their strength and weaknesses. Then we introduce ReLink, a novel Open IE system for extracting binary relations from open-domain text. Its working model is based on identifying correct phrases and linking them in the most proper way to reflect their relationship in a sentence. After establishing connections, it can easily extract relations by using several pre-defined patterns. Despite using only shallow linguistic features for extraction, it does not have the same weakness that existed in older systems, and it can also avoid many similar issues arising in recent Open IE systems. Our experiments show that ReLink achieves larger Area Under Precision-Recall Curve compared with ReVerb and Ollie, two well-known Open IE systems.