Compare Two Embedding Models — total

Given two alternative embeddings of a set of tokens or documents, total_dist() computes a global metric of the distance between the alternatives (by default the Wasserstein distance).

Usage

total_dist(
  x,
  y,
  matching = NULL,
  method = c("euclidean", "minkowski", "cosine", "cosine_squished", "dot_prod"),
  average = FALSE,
  ...
)

average_sim(x, y, matching = NULL, method = "cosine", average = TRUE, ...)

Arguments

x: an embeddings object
y: an embeddings object. If matching = NULL, y must contain at least a few rownames matching those of x.
matching: (optional) a named character vector specifying a one-to-one matching between rownames of x (names) and rownames of y (values)
method: either the name of a method to compute similarity or distance, or a function that takes two vectors, x and y, and outputs a scalar, similar to those listed in Similarity and Distance Metrics
average: logical. Should the rowwise distances be averaged as opposed to summed?
...: additional parameters to be passed to method function

Details

total_dist() computes the distance or similarity between the embeddings in x and their direct counterparts in y. average = TRUE returns the mean of these values, while average = FALSE (the default) returns the sum.

For more information on available methods, see get_sims().

method = "euclidean" and average = FALSE (the default for total_dist()) results in the Wasserstein distance, if embeddings are taken as equally weighted point masses.

average_sim() is identical to total_dist(), but with different defaults.