Distant Supervised Relation Extraction with Wikipedia and Freebase

In this paper we discuss a new approach to extract relational data from unstructured text without the need of hand labeled data. Socalled distant supervision has the advantage that it scales large amounts of web data and therefore fulfills the requirement of current information extraction tasks. As opposed to supervised machine learning we train generic, relationand domain-independent extractors on the basis of data base entries. We use Freebase as a source of relational data and a Wikipedia corpus tagged with unsupervised word classes. In contrast to previous work in the field of distant supervision, we do not rely on preprocessing steps that involve supervised learning. This work consists of three parts, a distant supervised Named Entity Recognizer (NER), a distant supervised classifier to recognize sentences in which a certain relation between two objects is described and the combination of both, allowing us for example to contribute new instances to Freebase. The performance of the NER is too low, that the combined method produces usable results. Still the subcomponents can be used independently.