Fit a multinomial or Bernoulli Naive Bayes model, given a dfm and some training labels.
textmodel_nb(x, y, smooth = 1, prior = c("uniform", "docfreq", "termfreq"), distribution = c("multinomial", "Bernoulli"))
the dfm on which the model will be fit. Does not need to contain only the training documents.
vector of training labels associated with each document identified
smoothing parameter for feature counts by class
prior distribution on texts; one of
count model for text features, can be
textmodel_nb() returns a list consisting of the following (where
\(I\) is the total number of documents, \(J\) is the total number of
features, and \(k\) is the total number of training classes):
original function call
\(k \times J\); probability of the word given the class (empirical likelihood)
\(k\)-length named numeric vector of class prior probabilities
\(k \times J\); posterior class probability given the word
\(J \times 1\); baseline probability of the word
the \(I \times J\) training dfm
y training class vector
the distribution argument
the prior argument
the value of the smoothing parameter
Prior distributions refer to the prior probabilities assigned to the training classes, and the choice of prior distribution affects the calculation of the fitted probabilities. The default is uniform priors, which sets the unconditional probability of observing the one class to be the same as observing any other class.
"Document frequency" means that the class priors will be taken from the relative proportions of the class documents used in the training set. This approach is so common that it is assumed in many examples, such as the worked example from Manning, Raghavan, and Schütze (2008) below. It is not the default in quanteda, however, since there may be nothing informative in the relative numbers of documents used to train a classifier other than the relative availability of the documents. When training classes are balanced in their number of documents (usually advisable), however, then the empirically computed "docfreq" would be equivalent to "uniform" priors.
prior to "termfreq" makes the priors equal to the proportions
of total feature counts found in the grouped documents in each training
class, so that the classes with the largest number of features are assigned
the largest priors. If the total count of features in each training class was
the same, then "uniform" and "termfreq" would be the same.
Manning, Christopher D., Prabhakar Raghavan, and Hinrich Schütze. 2008. An Introduction to Information Retrieval. Cambridge University Press (Chapter 13). https://nlp.stanford.edu/IR-book/pdf/irbookonlinereading.pdf
Jurafsky, Daniel, and James H. Martin. 2018. Speech and Language Processing. An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition. Draft of 3rd edition, September 23, 2018 (Chapter 4). https://web.stanford.edu/~jurafsky/slp3/4.pdf
## Example from 13.1 of _An Introduction to Information Retrieval_ txt <- c(d1 = "Chinese Beijing Chinese", d2 = "Chinese Chinese Shanghai", d3 = "Chinese Macao", d4 = "Tokyo Japan Chinese", d5 = "Chinese Chinese Chinese Tokyo Japan") trainingset <- dfm(txt, tolower = FALSE) trainingclass <- factor(c("Y", "Y", "Y", "N", NA), ordered = TRUE) ## replicate IIR p261 prediction for test set (document 5) (nb <- textmodel_nb(trainingset, y = trainingclass, prior = "docfreq"))#> #> Call: #> textmodel_nb.dfm(x = trainingset, y = trainingclass, prior = "docfreq") #> #> Distribution: multinomial; prior: docfreq; smoothing value: 1; 4 training documents; 6 fitted features.summary(nb)#> #> Call: #> textmodel_nb.dfm(x = trainingset, y = trainingclass, prior = "docfreq") #> #> Class Priors: #> (showing first 2 elements) #> N Y #> 0.25 0.75 #> #> Estimated Feature Scores: #> Chinese Beijing Shanghai Macao Tokyo Japan #> N 0.1474 0.2059 0.2059 0.2059 0.5091 0.5091 #> Y 0.8526 0.7941 0.7941 0.7941 0.4909 0.4909coef(nb)#> classes #> features N Y #> Chinese 0.1473684 0.8526316 #> Beijing 0.2058824 0.7941176 #> Shanghai 0.2058824 0.7941176 #> Macao 0.2058824 0.7941176 #> Tokyo 0.5090909 0.4909091 #> Japan 0.5090909 0.4909091predict(nb)#> d1 d2 d3 d4 d5 #> Y Y Y N Y #> Levels: N Y# contrast with other priors predict(textmodel_nb(trainingset, y = trainingclass, prior = "uniform"))#> d1 d2 d3 d4 d5 #> Y Y Y N N #> Levels: N Ypredict(textmodel_nb(trainingset, y = trainingclass, prior = "termfreq"))#> d1 d2 d3 d4 d5 #> Y Y Y N Y #> Levels: N Y## replicate IIR p264 Bernoulli Naive Bayes nb_bern <- textmodel_nb(trainingset, y = trainingclass, distribution = "Bernoulli", prior = "docfreq") predict(nb_bern, newdata = trainingset[5, ])#> d5 #> N #> Levels: N Y