A Machine Learning Approach to Recipe Text Processing

We propose a machine learning approach to recipe text processing problem aiming at converting a recipe text to a work flow. In this paper, we focus on the natural language processing (NLP) such as word identification, named entity recognition, and s yntac- tic analysis to extract predicate-argument structures (tu ples of a ver- bal expression and its arguments) from a sentence in a recipe text. Predicate-argument structures are subgraphs of the work flo w of a recipe. We solve these problems by methods based on machine learning techniques. The recipe domain is, however, different from the gen- eral domain in which many language resources are available. And we have to adapt NLP systems to the recipe texts by preparing anno- tated data in the recipe domain. To reduce the cost of the adaptation, we adopt a pointwise framework allowing to train analyzers from partially annotated data. The experimental results showed that an adaptation works well for each NLP and with all the adaptations the accuracy of the entire system increased. We can conclude that more adaptation work helps develop an accurate recipe-text-to-flow system.