Many traditional relation extraction techniques require a large number of pre-defined schemas in order to extract relations from textual documents. In this paper, to avoid the need for pre-defined schemas, we employ the notion of universal schemas that is formed as a collection of patterns derived from Open Information Extraction as well as from relation schemas of pre-existing datasets. We then employ matrix factorization and collaborative filtering on such universal schemas for relation extraction. While previous systems have trained relations only for entities, we exploit advanced features from relation characteristics such as clause types and semantic topics for predicting new relation instances. This helps our proposed work to naturally predict any tuple of entities and relations regardless of whether they were seen at training time with direct or indirect access in their provenance. In our experiments, we show improved performance compared to the state-of-the-art.
展开▼
