(A) siamese network for aligning open information extraction relations to knowledge base relations

Abstract

Open Information Extraction (Open IE) is a system which produces an output in the form of triples from a large amount of text. A triple consists of two entities and a relation describing the relationship between those entities. Even though Open IE relations are more extensible than those used in a traditional Information Extraction system, the former are ambiguous and lack of semantics. Thus, we need a method to canonicalize Open IE relations. The existing approaches mostly use a clustering method to canonicalize triples without considering the information in an existing Knowledge Base (KB) consisting of triples with a predefined set of relations. Dutta et al. attempted to align Open IE relations with KB relations. However, they used a rule-based approach that requires human efforts to define the rules. Motivated by this problem, in this thesis, we attempt to canonicalize Open IE relations with KB relations by using a Siamese Network model. For training, we attempt to automatically generate a training dataset using a distant supervision approach rather than relying on a hand-labeled dataset. In the experiment, we show that our model performs better than the baselines.