Relation extraction using a handful of labeled examples, a web search engine and imitation learning

In this paper we describe an approach to knowledge base population (KBP) that, unlike previous work which uses relatively large existing knowledge bases (KBs) as supervision, uses only a small number of annotated instances for each relation. Furthermore, we assume neither a closed set of candidate answers or an existing named entity recognizer. This scenario arises commonly in real-life applications in which a new KB must be built.

With these considerations in mind, we develop a two-stage KBP system that is able to learn from around 30 examples per relation and a web search engine. The system consists of a candidate answer recognizer followed by a relation extractor which are trained jointly using imitation learning. 

We collect data for two relations (architect name and completion year of buildings) and show that it is possible to achieve good extraction performance. Furthermore, we demonstrate that the jointly learned two-stage system outperforms a single-stage system using the same features and classification learner.