What does it take to bake a cake? The RecipeRef corpus and anaphora resolution in procedural text

Findings of the Association for Computational Linguistics: ACL 2022, pages 3481 - 3495 May 22-27, 2022 c⃝2022 Association for Computational Linguistics What does it take to bake a cake? The RecipeRef corpus and anaphora resolution in procedural text Biaoyan Fang1, Timothy Baldwin3,1 and Karin Verspoor2,1 1The University of Melbourne, Australia 2RMIT University, Australia 3MBZUAI, Abu Dhabi biaoyanf@student.unimelb.edu.au {tbaldwin,karin.verspoor}@unimelb.edu.au Abstract Procedural text contains rich anaphoric phe- nomena, yet has not received much attention in NLP. To fill this gap, we investigate the textual properties of two types of procedural text, recipes and chemical patents, and gener- alize an anaphora annotation framework devel- oped for the chemical domain for modeling anaphoric phenomena in recipes. We apply this framework to annotate the RecipeRef corpus with both bridging and coreference relations. Through comparison to chemical patents, we show the complexity of anaphora resolution in recipes. We demonstrate empirically that transfer learning from the chemical domain im- proves resolution of anaphora in recipes, sug- gesting transferability of general procedural knowledge. 1 Introduction Anaphora resolution is a core component in in- formation extraction tasks (Poesio et al., 2016; Rösiger, 2019) and critical for various downstream natural language processing tasks, such as named entity recognition (Dai et al., 2019) and machine translation (Stanovsky et al., 2019). It consists of two primary anaphoric types, coreference (Ng, 2017; Clark and Manning, 2015) and bridging (Asher and Lascarides, 1998; Rösiger et al., 2018). Most anaphora corpora (Pradhan et al., 2012; Ghad- dar and Langlais, 2016; Poesio et al., 2008), how- ever, only focus on either coreference or bridging. To fill the gap in anaphora resolution, it is becom- ing increasingly important to have both types anno- tated. Current research on anaphora resolution is mostly based on declarative text (Pradhan et al., 2012; Ghaddar and Langlais, 2016; Rösiger, 2018a; Hou et al., 2018), such as news or dialogue. Proce- dural text, such as chemical patents or instruction manuals, has received limited attention despite be- ing critical for human knowledge (Yamakata et al., 2020). In turn, correct resolution of entities is the cornerstone of procedural text comprehension— resolution of anaphora in these texts is required to determine what action applies to which entity. We focus in this work on the procedural text type of recipes. As shown in Fig. 1, recipes have rich and complex anaphora phenomena. Here, the expression the biscuits appears several times in text; while each occurrence relates to the same biscuits concept, their state and semantic meaning vary. Our aim in this paper is to address anaphora res- olution in procedural text, especially for recipes, identifying anaphoric references and determining the relationships among the entities. We first in- vestigate the textual properties of procedural texts, i.e. chemical patents and recipes. We then adapt an existing anaphora annotation schema developed for chemical patents (Fang et al., 2021a,b) to recipes, and define four types of anaphora relationships, en- compassing coreference and bridging. We further create a dataset based on this schema and achieve high inter-annotator agreement with two annota- tors …

Content truncated to 500 words