An approach to automatic text production in electronic medical record systems

Abstract

The paper describes basic properties of a sentence generator which requires minimal input information. Input is a set of unstructured semantic concepts, and the generator produces sentences which are compatible with this set by utilizing information from a statistical language model. Output is filtered by a simple context-free grammar. The system is trained on text from electronic medical records, and it is able to produce well-formed sentences in cases involving simple medication prescriptions and symptom descriptions. Basic complexity aspects of the problem are described, and suggestions for efficient implemented generators which manage to produce sentences within acceptable time limits, despite the complexity of the approach, are presented in the final sections. © Springer-Verlag Berlin Heidelberg 2005.