Distant Supervision for Relation Extraction with Type Constraint

Abstract

Distant supervision has been proposed for extracting relation instances from unstructured corpus. However, it is based on the assumption that each sentence mentioning two related entities is an expression of the given relation, which is not always the case and in turn can lead to noisy patterns and thus hurt precision. To improve the accuracy, this paper proposes an approach by adding entity's information as extra type constraint into the distance supervision based framework to train a relation extractor. To achieve this, a way to align knowledge base entities to their text mentions, and extract each entity's information (features) from those text mentions are provided in first. Secondly, two ways of adding entity's information to the problem of relation extraction are explored: (1) a joint algorithm, which models entity's features and relation's features jointly, (2) a type-constraint algorithm, which uses such features to constrain the types of relation arguments. The experimental results demonstrate that our approach can significantly improve the accuracy on the interesting datasets.