DKGBuilder: An architecture for building a domain knowledge graph from scratch

In recent years, we have witnessed the technical advances in general knowledge graph construction. However, for a specific domain, harvesting precise and fine-grained knowledge is still difficult due to the long-tail property of entities and relations, together with the lack of high-quality, wide-coverage data sources. In this paper, a domain knowledge graph construction system DKGBuilder is presented. It utilizes a template-based approach to extract seed knowledge from semi-structured data. A word embedding based projection model is proposed to extract relations from text under the framework of distant supervision. We further employ an is-a relation classifier to learn a domain taxonomy using a bottom-up strategy. For demonstration, we construct a Chinese entertainment knowledge graph from Wikipedia to support several knowledge service functionalities, containing over 0.7M facts with 93.1% accuracy.