Embedding-based pairwise construct-redundancy screen — tf_embedding

For every unordered pair of constructs, embeds each definition with the supplied embedder and computes the cosine similarity (rounded to 6 decimals). The vectors of every compared pair must be of equal, nonzero length, or the pair is refused. Returns a data frame sorted by descending cosine then (a, b), flagging pairs at or above threshold for review. This assistive screen complements the deterministic lexical tf_redundancy_check(), and its results are only as reproducible as the supplied embedder.

Usage

tf_embedding_redundancy(theory, embedder, threshold = NULL)

Arguments

theory: A theory object (named list), e.g. from tf_read().
embedder: A function mapping a definition string to a numeric vector.
threshold: Cosine threshold for the "review" flag; defaults to the checklist's embedding_similarity_max (0.85). How high a cosine two unrelated definitions reach depends on the embedding model, so a threshold suited to one model may flag everything or nothing under another, and the value is best set for the model in use.

Value

A data frame with columns a, b, cosine, flag.

Examples

theory <- tf_theory("demo-1", "A demonstration theory") |>
  tf_add_construct("c_arousal", "Arousal", "bodily activation") |>
  tf_add_construct("c_threat", "Threat", "appraised danger")
# A toy deterministic embedder: bag-of-words counts over a fixed vocabulary.
vocab <- c("bodily", "activation", "appraised", "danger")
embedder <- function(def) {
  words <- strsplit(tolower(def), "\\s+")[[1]]
  vapply(vocab, function(w) sum(words == w), numeric(1))
}
tf_embedding_redundancy(theory, embedder)
#>           a        b cosine flag
#> 1 c_arousal c_threat      0   ok