Add information on reference documents to predict.textmodel_wordscores

[stefan-mueller]

As far as I see, the predict() function for textmodel_wordscores() only contains the document names and the estimated scores. As a result, based on the predict() output we do not which documents were originally reference texts.

This could be somewhat problematic if a user plots the estimated scores with textplot_scale1d(margin = c("documents")), but only wants the scores for the virgin texts (or wants to highlight which documents were reference texts with a different shape/colour).

To add information on the reference documents, it would probably sufficient to add the textscore/NA to the predict.textodel_wordscores function. Then we could also adjust textplot_scale1d() and add an option such as include_refscores = FALSE or highlight_refscores = TRUE.

ws <- textmodel_wordscores(data_dfm_lbgexample, c(seq(-1.5, 1.5, .75), NA))

summary(ws)

# information on reference scores
ws$y
# > ws$y
# [1] -1.50 -0.75  0.00  0.75  1.50    NA

str(predict(ws))
# > str(predict(ws))
# Classes 'predict.textmodel_wordscores', 'numeric'  Named num [1:6] -1.32 -7.40e-01 -8.67e-18 7.40e-01 1.32 ...
# ..- attr(*, "names")= chr [1:6] "R1" "R2" "R3" "R4" ...

[kbenoit]

Understand the point, but we spent a lot of time trying to make the predict() methods for textmodel objects behave as closely to (e.g.) predict.lm() as possible. That’s why they predict whatever you ask them to predict on.

Two solutions: a) add an argument to predict.textmodel_wordscores() to exclude reference texts; or b) use textmodel_affinity(), which is the newer, better alternative to wordscores. And hopefully to be published soon.

[stefan-mueller]

This should avoid cases like this example from our tutorial where the reference text scores "distort" the plotted results.