The Zen of Elas¶
ASReview is a research project coordinated by Rens van de Schoot (full Professor at the Department of Methodology & Statistics and ambassador of the focus area Applied Data Science at Utrecht University, The Netherlands), together with Jonathan de Bruin, Lead engineer of the ASReview project and working at the Information and Technology Services department at Utrecht University.
Our advisory board consists of machine learning expert Daniel Oberski, associate professor at Utrecht University’s Department of Methodology & Statistics, and the department of Biostatistics at the Julius Center, University Medical Center Utrecht), full professor Lars Tummers (Professor of Public Management and Behavior at Utrecht University), Ayoub Bagheri (NLP-expert at Utrecht University), Bianca Kramer (Open Science expert at the Utrecht University library), Jan de Boer (Information specialist at the Utrecht university library), Felix Weijdema (Systematic review specialist at the Utrecht University library), and Martijn Hutijs (UX-expert at the department Test and Quality Services at Utrecht University).
The Art-Work of ASReview was developed by Joukje Willemsen.
Moreover, many others helped the project, like researchers Gerbrich Ferdinands and Laura Hofstee, as well as many students like Yongchao Terry Ma, Sofie van den Brand, Sybren Hindriks, and Albert Harkema. Many thanks to all the contributors!
Principles of Elas¶
Elas is the mascotte of ASReview and your Electronic Learning Assistant who will guide you through the interactive process of making decisions using Artificial Intelligence in ASReview. Elas comes with some important principles:
Humans are the Oracle It’s the interaction between humans and machines which will take science a major leap forward. We believe a human should make the final decision about whether to mark a record as relevant/irrelevant (hence, is the Oracle), and the software merely orders the records on the relevance score as predicted by the model in each iteration of the active learning cycle.
Open & Transparent We are strong opponents of open science, and therefore ASReview code is free and openly available. We value your privacy and hence do not get to see any of your data (everything stays on your device). We do hope you believe like us in the FAIR data principles and publish your data, results, and project file on a data repository.
Unbiasedness We signed the DORA-declaration, and we only present text for unbiased decision making. So, when screening for example academic papers we only show titles and abstracts, and we do not present authors, or journal names. This way, you can focus on what is truly important (the content) and don’t get tempted to use irrelevant information.
AI-aided Interface Simplicity is the ultimate sophistication (Davinci) and, therefore, we keep the front-end as clean as possible. Boring, but efficient because the magic happens under the hood.
Garbage in garbage out We focus on the machine learning part of the pipeline and not on the preprocessing or postprocessing of the data (which reference managers are designed for). Be aware of the principle GIGO and check the quality of your data first. Don’t blame Elas if the performance is not as good as expected due to low quality input data.
The Case of Systematic Reviewing¶
With the emergence of online publishing, the number of scientific papers on any topic, e.g. COVID19, is skyrocketing. Simultaneously, the public press and social media also produce data by the second. All this textual data presents opportunities to scholars, but it also confronts them with new challenges. To summarize all this data, researchers write systematic reviews, providing essential, comprehensive overviews of relevant topics. To achieve this, they have to screen (tens of) thousands of studies by hand for inclusion in their overview. As truly relevant papers are very sparse (i.e., often <10%), this is an extremely imbalanced data problem. The process of finding these rare relevant papers is error prone and very time intensive.
The rapidly evolving field of machine learning (ML) has allowed the development of ML-aided pipelines that assist in finding relevant texts for such search tasks. A well-established approach to increase the efficiency of title and abstract screening is determining prioritization with active learning, which is very effective for systematic reviewing.
The goal of ASReview is to help scholars and practitioners to get an overview of the most relevant records for their work as efficiently as possible, while being transparent in the process. It is uses active learning, allows multiple ML-models, and ships with a benchmark mode which is especially useful for comparing and designing algorithms. Furthermore, it is intended to be easily extensible, allowing third parties to add modules that enhance the pipeline and can process any text (although we consider systematic reviewing as a very useful approach).