Aligning Actions Across Recipe Graphs

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6930–6942 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 6930 Aligning Actions Across Recipe Graphs Lucia Donatelli, Theresa Schmidt, Debanjali Biswas, Arne K¨ohn, Fangzhou Zhai, Alexander Koller Department of Language Science and Technology Saarland Informatics Campus Saarland University {donatelli,theresas,dbiswas,koehn,fzhai,koller} @coli.uni-saarland.de Abstract Recipe texts are an idiosyncratic form of in- structional language that pose unique chal- lenges for automatic understanding. One chal- lenge is that a cooking step in one recipe can be explained in another recipe in differ- ent words, at a different level of abstraction, or not at all. Previous work has annotated corre- spondences between recipe instructions at the sentence level, often glossing over important correspondences between cooking steps across recipes. We present a novel and fully-parsed English recipe corpus, ARA (Aligned Recipe Actions), which annotates correspondences be- tween individual actions across similar recipes with the goal of capturing information implicit for accurate recipe understanding. We rep- resent this information in the form of recipe graphs, and we train a neural model for pre- dicting correspondences on ARA. We find that substantial gains in accuracy can be obtained by taking fine-grained structural information about the recipes into account. 1 Introduction Cooking recipes are a type of instructional text that many people interact with in their everyday lives. A recipe explains step by step how to cook a cer- tain dish, describing the actions a chef needs to perform as well as the ingredients and intermediate products of the cooking process. However, recipes for the same dish often differ in which cooking actions they describe explicitly, how they describe them, and in which order.1 For instance, in the three recipes in Fig. 1, the overall process of as- sembling a batter and making waffles is explained with different levels of detail: recipe (a) explains the process with three distinct cooking actions (in bold); recipe (b) with eight distinct actions; and recipe (c) with nineteen actions. 1For now, we understand an action as synonymous to a semantic event. We use the former term for consistency with other recipe work. (a) Beat eggs. Mix in remaining ingredients. Cook on hot waffle iron. 2 (b) Preheat your waffle iron. In a large bowl, mix together the flour, salt, baking powder, and sugar. In another bowl, beat the eggs. Add the milk, butter, and vanilla to the eggs. Pour the liquid into the flour mixture and beat until blended. Ladle the batter into the waffle iron and cook until crisp and golden. 3 (c) Sift together in a large mixing bowl flour, baking pow- der, salt, and sugar. In a jug, measure out milk. Separate eggs, placing egg whites in the bowl of standing mixer. Add yolks and vanilla essence to milk and whisk together. Pour over the flour mixture and very gently stir until about combined. Stir in the melted butter and continue mix- ing very gently until combined. Beat egg whites until stiff and slowly fold into batter. …

Content truncated to 500 words