Modeling Biological Processes for Reading Comprehension

Modeling Biological Processes for Reading Comprehension Jonathan Berant∗, Vivek Srikumar∗, Pei-Chun Chen, Brad Huang and Christopher D. Manning Stanford University, Stanford Abby Vander Linden and Brittany Harding University of Washington, Seattle Peter Clark Allen Institute for Artificial Intelligence, Seattle Abstract Machine reading calls for programs that read and understand text, but most current work only attempts to extract facts from redundant web-scale corpora. In this pa- per, we focus on a new reading compre- hension task that requires complex reason- ing over a single document. The input is a paragraph describing a biological pro- cess, and the goal is to answer questions that require an understanding of the re- lations between entities and events in the process. To answer the questions, we first predict a rich structure representing the process in the paragraph. Then, we map the question to a formal query, which is executed against the predicted structure. We demonstrate that answering questions via predicted structures substantially im- proves accuracy over baselines that use shallower representations. 1 Introduction The goal of machine reading is to develop pro- grams that read text to learn about the world and make decisions based on accumulated knowl- edge. Work in this field has focused mostly on macro-reading, i.e., processing large text collec- tions and extracting knowledge bases of facts (Et- zioni et al., 2006; Carlson et al., 2010; Fader et al., 2011). Such methods rely on redundancy, and are thus suitable for answering common factoid ques- tions which have ample evidence in text (Fader et al., 2013). However, reading a single document (micro-reading) to answer comprehension ques- tions that require deep reasoning is currently be- yond the scope of state-of-the-art systems. In this paper, we introduce a task where given a paragraph describing a process, the goal is to ∗Both authors equally contributed to the paper. answer reading comprehension questions that test understanding of the underlying structure. In par- ticular, we consider processes in biology text- books such as this excerpt and the question that follows: “. . . Water is split, providing a source of elec- trons and protons (hydrogen ions, H+) and giv- ing off O2 as a by-product. Light absorbed by chlorophyll drives a transfer of the electrons and hydrogen ions from water to an acceptor called NADP+ . . . ” Q What can the splitting of water lead to? a Light absorption b Transfer of ions This excerpt describes a process in which a com- plex set of events and entities are related to one another. A system trying to answer this ques- tion must extract a rich structure spanning multi- ple sentences and reason that water splitting com- bined with light absorption leads to transfer of ions. Note that shallow methods, which rely on lexical overlap or text proximity, will fail. Indeed, both answers are covered by the paragraph and the wrong answer is closer in the text to the question. We propose a novel method that tackles this challenging problem (see Figure 1). First, we train …

Content truncated to 500 words