Process-Level Representation of Scientific Protocols with Interactive Annotation

Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics, pages 2190–2202 April 19 - 23, 2021. ©2021 Association for Computational Linguistics 2190 Process-Level Representation of Scientific Protocols with Interactive Annotation Ronen Tamari:˚ Fan Bai; Alan Ritter; Gabriel Stanovsky:‹ :The Hebrew University of Jerusalem ;Georgia Institute of Technology ‹Allen Institute for Artificial Intelligence {ronent,gabis}@cs.huji.ac.il {fan.bai,alan.ritter}@cc.gatech.edu Abstract We develop Process Execution Graphs (PEG), a document-level representation of real-world wet lab biochemistry protocols, addressing challenges such as cross-sentence relations, long-range coreference, grounding, and implicit arguments. We manually annotate PEGs in a corpus of complex lab protocols with a novel interactive textual simulator that keeps track of entity traits and semantic constraints during annotation. We use this data to develop graph-prediction models, finding them to be good at entity identification and local relation extraction, while our corpus facilitates further exploration of challenging long-range relations.1 1 Introduction There is a drive in recent years towards automating wet lab environments, where menial benchwork procedures such as pipetting, centrifuging, or incubation are software-controlled, and either executed by fully automatic lab equipment (Lee and Miles, 2018), or with a human-in-the-loop (Keller et al., 2019). These environments allow reliable and precise experiment reproducbility while relieving researchers from tedious and laborious work which is prone to human error (Bates et al., 2017; Prabhu and Urban, 2017). To achieve this, several programmatic formalisms are developed to describe an experiment as an executable program. For example, Autoprotocol (Lee and Miles, 2018) defines a mix predicate taking three arguments: mode, speed, and duration. ˚Work begun on an internship at the Allen Institute for Artificial Intelligence. 1 Our annotated corpus, simulator, annotation interface, interaction data, and models are available for use by the research community at https://textlabs.github. io/. Figure 1: We develop a scaffold (center) between sentence-level lab procedure representations (top) and low-level, lab-specific instructions (bottom). The Process Execution Graph (PEG) captures document- level relations between procedures (orange rounded nodes) and their arguments (blue rectangular nodes). A promising direction to leverage automatic wet-lab environments is a conversion from natural language protocols, written in expressive free-form language, to low-level instructions, ensuring a non- ambiguous, repeatable description of experiments. In this work, we focus on a crucial first step towards such conversion – the extraction and representation of the relations conveyed by the protocol in a formal graph structure, termed Process Execution Graphs (PEG), exemplified in Figure 1. PEGs capture both concrete, exact quantities (“30 minutes”), as well as vague instructions (“swirl gently”). A researcher can then port the PEG (either manually or automatically) to their specific lab equipment, e.g., specifying what constitutes a gentle swirl setting and adding missing arguments, such as the temperature of the 2191 Figure 2: Example interaction with our simulator, showing predicate grounding (“chill” is a temp_type operation) input assignment (“vial” is an argument of “chill”), validation (warning for a missing argument) and auto-complete driven by state-tracking, where only legal instructions in a given state are presented. incubation in Figure 1. Formally, PEGs are directed, …

Content truncated to 500 words