Automatic search strategy reformulation interface for systematic reviews


Oral session: Searching and information retrieval (2)


Wednesday 23 October 2019 - 16:00 to 17:30


All authors in correct order:

Scells H1, Zuccon G1, Koopman B2, Clark J3
1 The University of Queensland, Australia
2 CSIRO, Australia
3 Centre for Research in Evidence Based Practice, Bond University, Australia
Presenting author and contact person

Presenting author:

Harrisen Scells

Contact person:

Abstract text
Background: within the last decade the rise of digital publishing has become widespread, enabling publications to be edited and updated after the fact. In the medical domain, systematic reviews are one type of digital document that is often updated after initial publication. This is usually because new evidence has been discovered and must be re-synthesized into the existing review. A problem, however, is that the initial search strategy used to identify the originally relevant studies may not be sufficient in capturing new studies, or may capture too many irrelevant studies. This means that time and effort must be spent reformulating new or variant search strategies. While this problem may be particularly pronounced in 'living systematic reviews', the problem of finding all relevant studies while minimizing irrelevant studies for typical systematic reviews is also difficult. This overarching problem signifies a gap to be filled with a system for automatic search strategy reformulation.

Objectives: the development of an automatic, interactive search strategy reformulation tool that assists researchers in updating systematic reviews and to improve existing search strategies.

Methods: the system proposed uses a recognised and effective theoretical framework, which automatically generates search strategy reformulations and selects the most effective variation. In this work, a user interface (Figure 1) is developed with the goal to insert a human-in-the-loop to drive the selection of the most effective search strategy. This interface is capable of:
1) tracking the effectiveness of reformulations over time, allowing users to manage their reformulation history by backtracking and jumping to previous search strategies;
2) evaluating the effectiveness of reformulations using standard information retrieval measures (e.g. precision, recall, F-measure), and domain-specific evaluation measures (e.g. Work Saved) by loading in a validation set of studies; and
3) filtering out studies that have already been screened (also by loading separately) in order to only show new studies.

Results: the theoretical framework, on which the generation and selection of search strategy reformulations is based, is shown to significantly improve the effectiveness of existing queries. Queries are shown to increase in effectiveness upwards of 100% to 200% and beyond, depending on the automatic selection process and evaluation measure.

Conclusions: a human-in-the-loop for the selection of search strategy reformulation allows users to have fine-grained control over the reformulation process. Allowing humans to drive the selection process in this framework is a new and novel approach, which has not yet been attempted. Finally, automatically generating reformulations removes possible human bias and error, and reduces the time and effort required to update a review.