Manual screening

In the systematic review community, the standard method to sort scientific material (such as articles and reports) is via manual screening. In general, this screening process is split into several stages: first titles are screened and the relevant articles kept; then the abstracts are screened for retained articles to see which of those may still be useful; and finally full-text articles are screened. revtools currently supports title and abstract screening only.

Title screening

Title screening is achieved using the function screen_titles. You can run this function without any commands to launch the app in ‘empty’ mode, allowing you to load data via the ‘browse’ button; or with an object of class bibliography or data.frame. As with screen duplicates, you get the same app - but slightly different behaviour - depending on how you launch it:

screen_titles() # 1. standalone; load in data in the app
result <- screen_titles() # 2. the same, but save back to workspace on exit
screen_titles(data) # 3. launch the app using data from the workspace
result <- screen_titles(data) # 4. specify an object to return data to

Once you’ve loaded the app using screen_titles, you will see the ‘Data’ tab open in the dashboard:

These four options work as follows:

Import allows you to drag-and-drop a dataset directly in to the app. If you've loaded screen_titles with data, or you have already added some data, then dragging and dropping here will add the new dataset to the existing one (using merge_columns in the background).

Save Data Allows you to save your progress to a .rds or .csv file.

Clear Data Wipes all data from the app (with a warning first to prevent mistakes).

Exit App closes the app, but also invisibly returns a data.frame to the workspace.

The ‘Appearance’ tab doesn’t affect the data itself, but how it is displayed.

This tab has the following options:

Number of articles shown allows you to increase or decrease the amount of information on your screen.

Order citations by enables you to shuffle between the order of the original dataset; alphabetical order by title; and random (set with rnorm).

Hide identifying information is set to TRUE by default, and means only the titles are shown. If set to FALSE, author and journal names are also shown.

Once you have loaded the app and added some data, you will be able to manually select or deselect articles one at a time, or in groups using the ‘select all’ and ‘exclude all’ buttons. You can also navigate between pages using the arrow buttons, or choose ‘unknown’ for ambiguous articles. As you make decisions about which titles to include, the color of the text will update to match your selection.

Abstract screening

Abstract screening is achieved using the function screen_abstracts. Unlike screen_titles, screen_abstracts only shows data for one article at a time; but the two apps are otherwise identical, with the exception that screen_abstracts allows you to make notes for each article should you wish.

Screening with topic models

The main way to investigate the results of a topic model in revtools is via the command screen_topics. This function behaves similarly to the other apps in this package, but it has many more options.

Selecting data

If you load screen_topics without any data, it will initially look quite empty:

From here there are a series of stages before you can see the results of your topic model. The first thing you should do is import data via the ‘Import’ button. Once you have done this, you will be faced with some new options:

Show one point per allows you to select what will be displayed by screen_topics. By default this is 'label', which is simply an index giving a unique value for each entry in the dataset, usually corresponding to a single book or article However, you can use this option to plot other interesting information. If, for example, you'd like each point to represent one journal, or one year, then you set that here.

Select included variables shows you a series of checkboxes that correspond to the columns of the underlying data.frame. Selecting a checkbox means that the text in that column will be passed to the the topic model. If you want to run the topic model only on titles, for example, then select 'titles' here; whereas if you'd like to include titles, abstracts and keywords you can select all three.

Once you have selected your data you are ready to go; but you still don’t have a plot to work with. To do that you have to use the ‘Model’ tab.

Topic model options

The ‘Model’ tab has only a few options, but they strongly affect the outcome of the plot displayed by screen_topics, as well as how long the code will take to run. There are three choices:

Model Type currently only supports Latent Dirichlet Allocation (LDA; the default) or Correlated Topic Models (CTMs). From what I've read, CTMs generally fit better, but generate topics that can be harder to distinguish.

Number of topics is a key parameter. There is no 'correct' answer to what the optimal number of topics should be: smaller numbers will give a broad overview but may group unlike concepts; while larger numbers will be more precise but potentially too atomized. Selecting larger number of topics also increases the time taken for the model to calculate.

Number of iterations is ignored for CTMs, but for LDA affects how hard the algorithm looks for an optimal fit. Smaller values will calculate quicker, but will be less reliable. The default (10k iterations) has been chosen as a pragmatic tradeoff between these considerations, and should not be considered ideal.

Running the topic model

One you hit ‘Calculate Model’, screen_topics first constructs a Document Term Matrix (DTM) from the specified data, and then runs the function make_dtm. This function performs the following transformations:

  1. converts all text to lower case
  2. removes punctuation, 'stop words' (defaulting to revwords) and numbers
  3. performs stemming on all words (warning: requires the SnowballC package)
  4. removes all words with <3 letters, <5 appearances, or that are in <1% of documents
  5. replaces stemmed words with the most common 'full' version of that word in the corpus

Once the DTM has been constructed, screen_topics runs the specified topic model. This can take a very long time for large corpora, so don’t be surprised if you are waiting a while looking at this screen:

Interpreting the plot

Once the topic model has finished running, it generates a plot known as an ‘ordination’. The axes on this plot don’t have any implicit meaning, and are unlabelled. Instead, points that are closer together are interpreted as containing similar topics, and therefore similar words. If you hover over a point on this plot, it will show you the title of the paper in question, as well as the highest-weighted topic for that article:

If you click on an article, then further information on that article will be displayed, including an abstract if one is available. You will also be given the option to select or exclude that article, and to make notes:

If you would prefer to see information on topics, you can hover over the barplot on the right-hand side of the window. Here you can also select topics by clicking, and to select or exclude whole topics (note in this plot the sidebar has been minimized):

Finally, it is possible to investigate highly-weighted words in each topic by selecting ‘Show Words’ from the ‘Display’ tab:

This will show you the topics barplot, and invite you to select a topic to investigate further.

If you select a topic in the bar plot on the right of the screen, you will now be shown a barplot of highly-weighted words in that topic. You can then select and exclude words that you don’t want to contribute to your next topic model. Alternatively, you can search for words and exclude them manually.