A framework for recipe text interpretation

In this paper we describe a method for converting a recipe text into a meaning representation. The meaning representation is a flow graph, whose vertices are important word sequences in cooking (recipe named entity; NE) and edges denote relationships among them. Our methods consists of three parts: word segmentation (WS), recipe NE recognition (NER), and flow graph construction. The first two processes are based on machine learning and are adapted to recipe texts. The last process is based on huristic rules. As an evaluation we tested three processes on an annotated corpus. The results showed that WS and recipe NER achieved high accuracies and that flow graph construction is relatively difficult having a large room for improvement.