Address Recognition with Robust NLU Technology

An analysis component recognizing addresses on printed documents and identifying the respective sender or recipient is presented. The component’s kernel is a standard chart parser for feature-based context-free grammars. The parsing procedure is enhanced with the island-parsing strategy and the ability for partial parsing. The needs of the realworld application within document analysis of printed documents are met by several pre-processing steps. Additionally, a robust address matcher compares the parsing result with the entries of an address data base. Thus, understanding in this context is the identification of a database entry appropriate to a printed address. The described work has been implemented as a component within a large document analysis system. First evaluation results for 230 documents have recently been worked out.