Version 4.12.3.1 is a release focussed mostly on EPPI-Visualiser, which is our latest tool for publishing, visualising and exploring review data. Most notably, EPPI-Visualiser now supports 3D (evidence) maps, which can be pre-specified by the review authors, but can also be generated on the fly by visitors of the Visualiser app. On the EPPI-Reviewer side, the deduplication algorithm has been upgraded: it now fully supports deduplication for references written in non-Latin alphabets (Chinese, Japanes, Korean, Arabic, Cyrillic, Greek, etc.) and has been concurrently refined, improving the reliability of "similarity" scores.

EPPI-Visualiser: 3D Evidence Maps

Evidence (gap) Maps are becoming an extremely popular tool in the evidence synthesis community, and for good reason. EPPI-Reviewer already supports the creation and publication of interactive maps via the "EPPI-Mapper" tool, however, the maps it produces are based on data-snapshots (maps do not "update" automatically) and are (mostly) limited to 3 dimensions. By contrast, EPPI-Visualiser exclusively uses "real-time" data, and is therefore always up to date; it also allows visitors to explore review data freely without privileging one or the other specific way to access and organise the data. Overall we consider EPPI-Visualiser and EPPI-Mapper two "complementary" apps, each designed with different needs in mind; thus, we expect them to be often used "in parallel".

In this release, EPPI-Visualiser acquired the ability to present/visualise data in "map" form, meaning that it now supports "3D" tables: two dimensions are reported as rows and columns and the third dimension (called "segments") is represented within each resulting cell. Visitors of any given visualisation can now generate such 3D maps "on demand", using whichever branches of the published coding tools they may see fit. Concurrently, the data owners / review administrators can pre-specify "ready-made" maps, from the "Setup Visualisations" page in EPPI-Reviewer Web.

3D Maps in Visualiser offer two alternative views, a "bubble" view (default) where segments are shown as circles of different sizes (depending on how many references belong to each segment) or "table" view, that shows all numbers directly (and includes some more data as well). As per (1D) frequencies and (2D) crosstabs, maps allow to reach any list of references they represent, with a single click.

EPPI-Visualiser is able to plot between 1 and 6 segments, while EPPI-Mapper supports zero to four segments; moreover, Maps (and crosstabs) in EPPI-Visualiser can plot data comings from separate coding tools (each dimension can be selected from any coding tool, independently). Currently, Visualiser maps do not allow users to specify additional filters, although we expect to add this functionality in a later release. We also expect to add more ways to represent the segments (e.g. the mosaic view offered by EPPI-Mapper) at a later date.

EPPI-Visualiser: General Navigation

EPPI-Visualiser has also been upgraded in terms of navigation. Most pages will now "load" immediately and, if the data to be represented take a little while to "load" will show a "loading..." icon. Previously, some pages could take a few seconds to load, offering little visual clues that a little patience was required. Moreover, all pages are now able to "reuse" the data they are representing and thus generate alternative representations without having to reload the data from our servers.
Together, these changes produce a much improved "navigation" experience, where there are overall fewer "loading..." delays and those that are inevitable are better signposted. The layout of all pages was also improved, so to work a little better on small screens (although we are aware more improvements are needed to fully support small screens).

EPPI-Visualiser: new/improved visualisations

In the "frequencies results" page, a new data representation ("treemap") is present; furthermore, most existing representations (pie, bar chart) are now enhanced in their ability to represent data, adapting better to the number of codes, long code names and so forth. Finally, pie and bar charts can now be "saved" as images.

EPPI-Reviewer (both versions): deduplication algorithm

Until this release, the deduplication algorithm was performing poorly when deduplicating references written in non-Latin alphabets (Chinese, Japanese, Korean, Arabic, Cyrillic, Greek, etc.), moreover, having many hundreds of such references in a given review could prevent the "Find new duplicates" procedure from completing successfully. The present release resolves this problem and concurrently ensures that references written in non-Latin alphabets can be "grouped" sensibly and that their similarity scores do work as intended. We expect the accuracy of our algorithm to be roughly comparable between references that use the Latin alphabet and those that do not; however, it is likely that, depending on the relevant languages, the accuracy for the latter might be marginally degraded (although much better than what was possible until now).

Concurrently, the overall "scoring and grouping" algorithm has been upgraded and should now produce more reliable/informative similarity scores overall (please note: existing scores will not update automatically. To recalculate them on a per-group basis, you can change and reinstate the "master" item). The changes made in this respect are too complicated to be explained here, however, according to our tests, they significantly reduce scores for "false positives" and might consequently marginally increase the number of references that require manual checking (when the machine cannot possibly make accurate decisions).

EPPI-Reviewer Web: DOIs in duplicate groups

In the "duplicates" page, the DOI data for group members is now explicitly shown and colour-coded to highlight differences. The underlying algorithm does indeed use DOIs to calculate similarity scores (and grouping); in fact, it is built on the assumption that DOIs (when present) are highly significant. Thus, it is important to show this information, as it both helps "explaining" similarity scores and is potentially very useful if/when manual deduplication happens.

EPPI-Reviewer Web: bugfix

The previous release included a new "apply classifiers built in a different review" feature, which alas, wasn't actually working. We have resolved this problem and this feature is now fully functional.

Update: version 4.12.3.2

This release (27/10/2021) does not include visible new features, it is an "out of (normal) schedule" release to fix bugs, one of which was urgent.

Critical Bugfix: "Find new duplicates" could fail to complete (EPPI-Reviewer, both versions).
Version 4.12.3.1 significantly enhanced support for references written in non-Latin alphabets, making a new problem emerge: one of the routines that "cleans" text for the purpose of comparing specific fields from separate references could enter in an endless loop, as ordinary string manipulation commands appear to malfunction, when odd / rarely used Unicode characters appear in the input string. This bug, although extremely rare in practice, had high impact on the affected users (would make it impossible to update duplicate groups after importing more references) and EPPI-Reviewer performance (as it would use processing power for no gain, for several hours, if left unchecked). The problem is now resolved.

Bugfix: matching of references written in languages that are read from right to left.
EPPI-Reviewer uses an initial heuristic system to identify pairs of references that might be duplicates, and thus merit in-depth evaluation. This initial step relies on the text in the "Title" field only. When dealing with titles written in languages that are read from right to left, the code attempted to account for this difference, but wasn't working as intended, and was thus less reliable than the equivalent code that is used for the more frequent "left to right" languages. The problem is now resolved.

Marginal improvement: "Find new duplicates".
The routine we use to "clean text" (the same one mentioned above) works differently when dealing with non-Latin alphabets. The code to detect when this is the case was a little too sensitive and was triggered more often than planned. Extensive in-depth analysis showed that the effect was very small (just enough to be detected), but might marginally increase the number of false negatives. Since triggering the code that deals with non-Latin alphabets is also an additional computational cost, the updated version now triggers it less often, as per the original plan.

Bugfix: MAG pages, "bring review up to date" searches.
These searches could fail to consider one of the references used as the seeding data. The problem could have a real-world impact when using small batches of references for seeding and is now resolved.

New feature (not visible): EPPI-Visualiser now records some basic data about visits/usage.
This data however is currently only visible to the EPPI-Reviewer team, while we check if what is being recorded appears to be useful, sufficient and at the right level of detail. Once we'll confirm that the chosen logging strategy appears to be fit for purpose, we plan to make these stats available to review administrators in the "Setup Visualisation" page of EPPI-Reviewer Web.