Sentiment analysis with RNN or transformer on tokens

Orchestrates tokenise -> caller model -> softmax -> (optional) counted accuracy/confusion/precision/recall/F1.

Usage

morie_geron_sentiment_analysis(
  texts,
  model,
  tokenizer = NULL,
  y_true = NULL,
  labels = NULL
)

Arguments

texts

Character vector of documents.

model

Function(tokens) -> numeric scores, length K >= 2, constant across texts.

tokenizer

Optional function(text) -> character vector of tokens; default lowercase+split.

y_true

Optional integer gold labels in 0..K-1.

labels

Optional character vector of class names, length K.

Value

list with probabilities, predicted, tokens, labels, accuracy, confusion, precision, recall, f1, macro_f1, n_classes, estimate, n, method.

Examples

model <- function(toks) {
  s <- sum(toks %in% c("good", "great")) - sum(toks %in% c("bad", "awful"))
  c(-s, s)
}
texts <- c("a good great film", "an awful bad mess")
r <- morie_geron_sentiment_analysis(texts, model, y_true = c(1L, 0L))
str(r, max.level = 1)
#> List of 14
#>  $ probabilities: num [1:2, 1:2] 0.018 0.982 0.982 0.018
#>  $ predicted    : int [1:2] 1 0
#>  $ tokens       :List of 2
#>  $ labels       : NULL
#>  $ accuracy     : num 1
#>  $ confusion    : int [1:2, 1:2] 1 0 0 1
#>  $ precision    : num [1:2] 1 1
#>  $ recall       : num [1:2] 1 1
#>  $ f1           : num [1:2] 1 1
#>  $ macro_f1     : num 1
#>  $ n_classes    : int 2
#>  $ estimate     : num 1
#>  $ n            : int 2
#>  $ method       : chr "Tokenise -> model scores -> softmax, with counted accuracy / confusion / macro-F1"