A Hierarchical Model for Universal Schema Relation Extraction

Relation extraction by universal schema avoids mapping to a brittle, incomplete traditional schema by instead making predictions in the union of all input schemas, including textual patterns. Modeling these predictions by matrix competition with matrix factorization has yielded state-of-the-art accuracies. One difficulty with prior work in matrix factorization, however, is that there is no negative training data. As a result, existing methods merely sample an entity-pair’s unobserved relation types and assume they are negative. In this paper we instead maximize the likelihood of the observed data—achieving tractability by arranging the relation types as leaves in a binary tree. We show empirically that the choice of tree structure is consequential, and achieve a 2.61% F1 score improvement over the previous approach. Furthermore, a simple combination of this approach with the previous approach results in a 3.53% gain in F1 score.