A Robust Number Parser Based on Conditional Random Fields

Abstract

When processing information from unstructured sources, numbers have to be parsed in many cases to do useful reasoning on that information. However, since numbers can be expressed in different ways, a robust number parser that can cope with number representations in different shapes is required in those cases. In this paper, we show how to train such a parser based on Conditional Random Fields. As training data, we use pairs of Wikipedia infobox entries and numbers from public knowledge graphs. We show that it is possible to parse numbers at an accuracy of more than 90%.