Predicting the Structure of Cooking Recipes

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 781–786, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Predicting the Structure of Cooking Recipes Jermsak Jermsurawong and Nizar Habash Computational Approaches to Modeling Language Lab Computer Science New York University Abu Dhabi United Arab Emirates {jermsak.jermsurawong,nizar.habash}@nyu.edu Abstract Cooking recipes exist in abundance; but due to their unstructured text format, they are hard to study quantitatively beyond treating them as simple bags of words. In this paper, we propose an ingredient- instruction dependency tree data structure to represent recipes. The proposed rep- resentation allows for more refined com- parison of recipes and recipe-parts, and is a step towards semantic representation of recipes. Furthermore, we build a parser that maps recipes into the proposed rep- resentation. The parser’s edge prediction accuracy of 93.5% improves over a strong baseline of 85.7% (54.5% error reduction). 1 Introduction Cooking recipes are a specific genre of how-to in- structions which have been gaining interest in re- cent years as they may allow us to discover in- sights into culinary and cultural preferences. Most of the work studying recipes relies on simple in- gredient bag-of-word representations. While such representations may suffice for many purposes, they fail to capture much of the recipes’ internal structure. A smaller number of efforts focus on se- mantic representations of cooking recipes with an eye toward more complex and deep understand- ing. Natural language processing (NLP) tech- niques have been used to interpret cooking instruc- tions; however, the results have not been as suc- cessful as other language genres due to the unique aspects of cooking recipes. This paper makes two contributions. First, we propose an ingredient-instruction dependency tree representation of recipe structure. This represen- tation abstracts textual recipes more expressively than bag-of-word representations; and it allows for more nuanced comparisons of recipes. Second, we present a cooking recipe parser that maps text recipes into our proposed ingredient-instruction tree structures. The overall accuracy of predicting edges of our ingredient-instruction trees is 93.5%, beating a strong baseline of 85.7%, and achieving a relative error reduction of 54.5%. We present next some related work (Section 2) followed by a discussion of the structure of recipes and our representation (Section 3). Section 4 de- tails the recipe parser design, implementation and evaluation. 2 Related Work There have been many efforts on the processing of cooking recipes using models that range from bags of words to complex semantic representations. Among the approaches to studying recipes as ingredient bags of words, Ahn et al. (2011) con- structed a data-driven flavor network relating in- gredients together. Jain et al. (2015) adopted Ahn et al. (2011)’s framework to further analyze culi- nary practices of specific cultures. Nedovic (2013) examined underlying ingredient groupings from their recipe co-occurrences, using topic modeling techniques (latent Dirichlet allocation), and fur- ther improvised novel ingredient combinations us- ing deep belief networks. Among the structured representation ap- proaches, Tasse and Smith (2008) proposed MILK (Minimal Instruction Language for the Kitchen), …

Content truncated to 500 words