Reasoning over Entity-Action-Location Graph for Procedural Text Understanding

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 5100–5109 August 1–6, 2021. ©2021 Association for Computational Linguistics 5100 Reasoning over Entity-Action-Location Graph for Procedural Text Understanding Hao Huang1∗, Xiubo Geng2†, Pei Jian3, Guodong Long1, Daxin Jiang2† 1Australian AI Institute, School of CS, FEIT, University of Technology Sydney 2STCA NLP Group, Microsoft 3School of Computing Science, Simon Fraser University {hao.huang-4,guodong.long}@{student.uts,uts}.edu.au {xigeng,djiang}@microsoft.com, jpei@cs.sfu.ca Abstract Procedural text understanding aims at track- ing the states (e.g., create, move, destroy) and locations of the entities mentioned in a given paragraph. To effectively track the states and locations, it is essential to capture the rich semantic relations between entities, ac- tions, and locations in the paragraph. Al- though recent works have achieved substan- tial progress, most of them focus on leverag- ing the inherent constraints or incorporating external knowledge for state prediction. The rich semantic relations in the given paragraph are largely overlooked. In this paper, we pro- pose a novel approach (REAL) to procedural text understanding, where we build a general framework to systematically model the entity- entity, entity-action, and entity-location rela- tions using a graph neural network. We fur- ther develop algorithms for graph construc- tion, representation learning, and state and lo- cation tracking. We evaluate the proposed ap- proach on two benchmark datasets, ProPara, and Recipes. The experimental results show that our method outperforms strong baselines by a large margin, i.e., 5.0% on ProPara and 3.2% on Recipes, illustrating the utility of se- mantic relations and the effectiveness of the graph-based reasoning model. 1 Introduction Procedural text often consists of a sequence of sen- tences describing processes, such as a phenomenon in nature (e.g., how sedimentary rock forms) (Dalvi et al., 2018) or instructions to complete a task (e.g., the recipe of Mac and Cheese) (Bosselut et al., 2018). Given a paragraph and its partic- ipant entities, the task of procedural text under- standing is to track the states (e.g., create, move, destroy) and locations (a span in the text) of the ∗Work is done during internship at Microsoft. †Corresponding author. entities. Compared with traditional machine read- ing task, which mainly focuses on the static rela- tions among entities, procedural text understanding is more challenging since it involves discovering complex temporal-spatial relations among various entities from the process dynamics. To effectively track the states and locations of entities, it is crucial to systematically model rich relations among various concepts in the paragraph, including entities, actions, and locations. Three types of relations are of particular interest. First, mentions of the same entity in different sentences are related. The inherent relation among these mentions may provide clues for a model to generate consistent predictions about the entity. For example, the entity electrical pulses are mentioned in two sentences “The retina’s rods and cones con- vert it to electrical pulses. The optic nerve carries electrical pulses through the optic canal.”. Con- necting its two mentions in two sentences helps to infer …

Content truncated to 500 words