Understanding Procedural Text using Interactive Entity Networks

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 7281–7290, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 7281 Understanding Procedural Text using Interactive Entity Networks Jizhi Tang, Yansong Feng, Dongyan Zhao Wangxuan Institute of Computer Technology, Peking University The MOE Key Laboratory of Computational Linguistics, Peking University {tangjizhi, fengyansong, zhaody}@pku.edu.cn Abstract The task of procedural text comprehension aims to understand the dynamic nature of en- tities/objects in a process. Here, the key is to track how the entities interact with each other and how their states are changing along the procedure. Recent efforts have made great progress to track multiple entities in a procedu- ral text, but usually treat each entity separately and ignore the fact that there are often multi- ple entities interacting with each other during one process, some of which are even explic- itly mentioned. In this paper, we propose a novel Interactive Entity Network (IEN), which is a recurrent network with memory equipped cells for state tracking. In each IEN cell, we maintain different attention matrices through specific memories to model different types of entity interactions. Importantly, we can update these memories in a sequential manner so as to explore the causal relationship between entity actions and subsequent state changes. We eval- uate our model on a benchmark dataset, and the results show that IEN outperforms state- of-the-art models by precisely capturing the interactions of multiple entities and explicitly leverage the relationship between entity inter- actions and subsequent state changes. Our code is available at: https://github.com/ esddse/IEN. 1 Introduction Procedural texts, e.g., scientific articles, instruction books, or recipes, are widely spread and useful. En- tity state tracking is the key task for procedural text comprehension. Usually, an entity could have sev- eral targets (e.g., existence, location) to be tracked, and the system needs to predict the target changes of each entity involved in the process. State track- ing is challenging because of the dynamic nature, the involvement of multiple entities, and the com- plexity of tracking targets. Most recent approaches often use an RNN-based method to model the state changes across the pro- cess in an entity-by-entity manner, and use different classifiers for different targets (Dalvi et al., 2018; Tandon et al., 2018; Gupta and Durrett, 2019). While these models can learn to leverage either local or global information and make state predic- tions to one entity with fair accuracy, they ignore the fact that there are often multiple entities inter- acting with each other during the procedure. Here is an example that describes a chemical transforma- tion step in the photosynthesis process: The water breaks into oxygen, hydrogen, and electrons. To perfectly capture this chemical conversion in the process, ”entity-by-entity” systems need to sep- arately find out that water is destroyed at this step, oxygen is created, hydrogen is created, and elec- trons are created at the same step, respectively. It is easy to see that if the system ignores the inter- actions among multiple entities, it will be prone …

Content truncated to 500 words