Recipe Instruction Semantics Corpus (RISeC): Resolving Semantic Structure and Zero Anaphora in Recipes

Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 821–826 December 4 - 7, 2020. c⃝2020 Association for Computational Linguistics 821 Recipe Instruction Semantics Corpus (RISeC): Resolving Semantic Structure and Zero Anaphora in Recipes Yiwei Jiang, Klim Zaporojets, Johannes Deleu, Thomas Demeester, Chris Develder Ghent University – imec, IDLab Ghent, Belgium {first name.last name}@ugent.be Abstract We propose a newly annotated dataset for information extraction on recipes. Unlike previous approaches to machine comprehension of procedural texts, we avoid a priori pre-defining domain-specific predicates to recognize (e.g., the primitive instructions in MILK) and focus on basic understanding of the expressed semantics rather than directly reduce them to a simplified state representation (e.g., ProPara). We thus frame the semantic comprehension of procedural text such as recipes, as fairly generic NLP subtasks, covering (i) entity recognition (ingredients, tools and actions), (ii) relation extraction (what ingredients and tools are involved in the actions), and (iii) zero anaphora resolution (link actions to implicit arguments, e.g., results from previous recipe steps). Further, our Recipe Instruction Semantic Corpus (RISeC) dataset includes textual descriptions for the zero anaphora, to facilitate language generation thereof. Besides the dataset itself, we contribute a pipeline neural architecture that addresses entity and relation extraction as well as identification of zero anaphora. These basic building blocks can facilitate more advanced downstream applications (e.g., question answering, conversational agents). 1 Introduction Recently, several efforts have aimed at understanding recipe instructions (see Section 2). We consider such recipes as prototypical for procedural texts, for which processing is complex due to the need to (i) understand the ordering of steps (not unlike, e.g., event ordering in news), (ii) solve frequent ellipsis (i.e., zero anaphora) and coreference resolution, and (iii) track the state changes they involve (e.g., ingredients processed/combined to new entities). Especially the latter distinguishes procedural text processing from more traditional information extraction (e.g., from news). Most existing works on recipes focus on recognizing pre-defined predicates, typically in the form of a limited set of instruction types (e.g., to convert the recipe to robot instructions) with predefined argument slots to fill. Further, they often rely on an available starting list of ingredients (which may not be available in other procedural text). Hence, current approaches towards recipe understanding make assumptions that are rather domain specific. In contrast, we aim for a more basic and generic structured representation of the procedural text, limiting domain-specific knowledge and building on more general semantic concepts. In particular, we build on semantic concepts as defined in PropBank (Kingsbury and Palmer, 2002), which are not domain-specific. Note that our proposed form of structured representations not necessarily allows directly solving informational queries that require explicit reasoning and/or state tracking (e.g., “Where are the tomatoes after step 5?”). We however pose that properly detecting the various entities (e.g., ingredients and their derivations) and the actions that are executed on them (as described by verbs), with the appropriate coreference and zero anaphora …

Content truncated to 500 words