Show Me More Details: Discovering Hierarchies of Procedures from Semi-structured Web Data

Show Me More Details: Discovering Hierarchies of Procedures from Semi-structured Web Data Shuyan Zhou♠∗, Li Zhang♣∗, Yue Yang♣, Qing Lyu♣, Pengcheng Yin♠, Chris Callison-Burch♣, Graham Neubig♠ ♠Carnegie Mellon University ♣University of Pennsylvania {shuyanzh,pcyin,gneubig}@cs.cmu.edu {zharry,yueyang1,lyuqing,ccb}@seas.upenn.edu Abstract Procedures are inherently hierarchical. To make videos, one may need to purchase a cam- era, which in turn may require one to set a budget. While such hierarchical knowledge is critical for reasoning about complex proce- dures, most existing work has treated proce- dures as shallow structures without modeling the parent-child relation. In this work, we at- tempt to construct an open-domain hierarchi- cal knowledge-base (KB) of procedures based on wikiHow, a website containing more than 110k instructional articles, each documenting the steps to carry out a complex procedure. To this end, we develop a simple and effi- cient method that links steps (e.g., purchase a camera) in an article to other articles with similar goals (e.g., how to choose a camera), recursively constructing the KB. Our method significantly outperforms several strong base- lines according to automatic evaluation, hu- man judgment, and application to downstream tasks such as instructional video retrieval.1 1 Introduction A procedure includes some steps needed to achieve a particular goal (Momouchi, 1980). Procedures are inherently hierarchical: a high-level procedure is composed of many lower-level procedures. For example, a procedure with the goal make videos consists of steps like purchase a camera, set up lighting, edit the video, and so on, where each step itself is a procedure as well. Such hierarchical relations between procedures are recursive: the lower-level procedures can be further decomposed into even more fine-grained steps: one may need to arrange the footage in order to edit the video. Relatively little attention has been paid to hierar- chical relations in complex procedures in the field ∗Equal contribution. 1A demo with partial data can be found at https://wikihow-hierarchy.github.io/. The code and the data are at https://github.com/shuyanzhou/wikihow_hierarchy. of NLP. Some work performed a shallow one-level decomposition and often required costly resources such as human expert task-specific annotation (Chu et al., 2017; Zhang et al., 2020a, 2021). More at- tention has been paid in fields adjacent to NLP. For example, Lagos et al. (2017) and Pareti et al. (2014) both create hierarchical structures in how-to docu- ments by linking action phrases in one procedure to another procedure or by linking steps in how- to articles to resources like DBPedia (Auer et al., 2007). This kind of linking is helpful for explain- ing complex steps to readers who do not have prior knowledge of the topic being explained. In this paper, we revisit this important but un- derstudied task to develop a simple and effective algorithm (Figure 1) to construct a hierarchical knowledge-base (KB) for over 110k complex pro- cedures spanning a wide range of topics from wiki- How, a large-scale how-to website that has recently become a widely-used resource in NLP (Zhou et al., 2019; Zellers et al., 2019; Zhang et al., 2020d,c).2 From each wikiHow article which represents a pro- cedure, …

Content truncated to 500 words