Everything Happens for a Reason: Discovering the Purpose of Actions in Procedural Text

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 4496–4505, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 4496 Everything Happens for a Reason: Discovering the Purpose of Actions in Procedural Text Bhavana Dalvi Mishra*, Niket Tandon∗, Antoine Bosselut, Wen-tau Yih, † Peter Clark Allen Institute for Artificial Intelligence, Seattle, WA {bhavanad, nikett, antoineb, peterc}@allenai.org Abstract Our goal is to better comprehend procedu- ral text, e.g., a paragraph about photosynthe- sis, by not only predicting what happens, but why some actions need to happen before oth- ers. Our approach builds on a prior process comprehension framework for predicting ac- tions’ effects, to also identify subsequent steps that those effects enable. We present our new model (XPAD) that biases effect predictions towards those that (1) explain more of the ac- tions in the paragraph and (2) are more plau- sible with respect to background knowledge. We also extend an existing benchmark dataset for procedural text comprehension, ProPara, by adding the new task of explaining actions by predicting their dependencies. We find that XPAD significantly outperforms prior systems on this task, while maintaining the perfor- mance on the original task in ProPara. The dataset is available at http://data.allenai. org/propara 1 Introduction Procedural text is common in natural language, for example in recipes, how-to guides, and science pro- cesses, but understanding it remains a major chal- lenge. While there has been substantial prior work on extracting the sequence of actions (“scripts”) from such texts, the task of identifying why the sequence is the way it is has received less attention, and is the goal of this work. While “why” can mean many things, we treat it here as describing how one action produces effects that are required by another. For example in Figure 1, “CO2 enters the leaf” is necessary because it results in “CO2 is at the leaf”, a precondition for “CO2 forms sugar”. (Note that ∗*Bhavana Dalvi Mishra and Niket Tandon contributed equally to this work. 1Scott is currently at Facebook AI Research (scot- tyih@fb.com) Figure 1: An action sequence (black) describes the or- der that actions happen, but not why. We add the new task of also predicting why the actions are needed, in the form of the actions’ effects (blue) and subsequent actions that depend on those effects (green). While a system that predicts effects without considering depen- dencies can make errors (red, top part), we obtain better predictions by biasing the system to predict effects that also result in more dependencies (lower part). an action does not directly depend on previous ac- tions, but rather on the state of the world result- ing from previous actions). If one could determine such rationales, new capabilities would become possible, including explanation, identifying alter- native event orderings, and answering "what if..." questions. However, this task is challenging as it requires knowledge of the preconditions and effects of actions, typically unstated in the text itself. …

Content truncated to 500 words