Nothing
R/modelling.R
In crfsuite: Conditional Random Fields for Labelling Sequential Data in Natural Language Processing
Defines functions
crf_options
predict.crf
summary.crf
print.crf
as.crf
crf
Documented in
as.crf
crf
crf_options
predict.crf
#' @title Linear-chain Conditional Random Field
#' @description Fits a Linear-chain (first-order Markov) CRF on the provided label sequence and saves it on disk in order to do sequence labelling.
#' @param x a character matrix of data containing attributes about the label sequence \code{y} or an object which can be coerced to a character matrix.
#' It is important to note that an attribute which has the same value in a different column is considered the same.
#' @param y a character vector with the sequence of labels to model
#' @param group an integer or character vector of the same length as \code{y} indicating the group the sequence \code{y} belongs to (e.g. a document or sentence identifier)
#' @param method character string with the type of training method. Either one of:
#' \itemize{
#' \item{lbfgs: }{L-BFGS with L1/L2 regularization}
#' \item{l2sgd: }{SGD with L2-regularization}
#' \item{averaged-perceptron: }{Averaged Perceptron}
#' \item{passive-aggressive: }{Passive Aggressive}
#' \item{arow: }{Adaptive Regularization of Weights (AROW)}
#' }
#' @param options a list of options to provide to the training algorithm. See \code{\link{crf_options}} for possible options and the example below on how to provide them.
#' @param embeddings a matrix with the same number of rows as \code{x} and in the same order with numeric information used in model building (experimental)
#' @param file a character string with the path to the file on disk where the CRF model will be stored.
#' @param trace a logical indicating to show the trace of the training output. Defaults to \code{FALSE}.
#' @param FUN a function which can be applied on raw text in order to obtain the attribute matrix used in \code{predict.crf}. Currently not used yet.
#' @param ... arguments to FUN. Currently not used yet.
#' @return an object of class crf which is a list with elements
#' \itemize{
#' \item{method: }{The training method}
#' \item{type: }{The type of graphical model which is always set crf1d: Linear-chain (first-order Markov) CRF}
#' \item{labels: }{The training labels}
#' \item{options: }{A data.frame with the training options provided to the algorithm}
#' \item{file_model: }{The path where the CRF model is stored}
#' \item{attribute_names: }{The column names of \code{x}}
#' \item{log: }{The training log of the algorithm}
#' \item{FUN: }{The argument passed on to FUN}
#' \item{ldots: }{A list with the arguments passed on to ...}
#' }
#' @references More details about this model is available at \url{http://www.chokkan.org/software/crfsuite/}.
#' @export
#' @seealso \code{\link{predict.crf}}
#' @examples
#' ## Download modeldata (conll 2002 shared task in Dutch)
#' \donttest{
#' x <- ner_download_modeldata("conll2002-nl")
#' }
#' # for CRAN only - word on a subset of the data
#' x <- ner_download_modeldata("conll2002-nl", docs = 10)
#' if(is.data.frame(x)){
#' ##
#' ## Build Named Entity Recognition model on conll2002-nl
#' ##
#' x$pos <- txt_sprintf("Parts of Speech: %s", x$pos)
#' x$token <- txt_sprintf("Token: %s", x$token)
#' crf_train <- subset(x, data == "ned.train")
#' crf_test <- subset(x, data == "testa")
#'
#' model <- crf(y = crf_train$label,
#' x = crf_train[, c("token", "pos")],
#' group = crf_train$doc_id,
#' method = "lbfgs",
#' options = list(max_iterations = 3, feature.minfreq = 5,
#' c1 = 0, c2 = 1))
#' model
#' weights <- coefficients(model)
#' head(weights$states, n = 20)
#' head(weights$transitions, n = 20)
#' stats <- summary(model, "modeldetails.txt")
#' stats
#' plot(stats$iterations$loss)
#'
#' ## Use the CRF model to label a sequence
#' scores <- predict(model,
#' newdata = crf_test[, c("token", "pos")],
#' group = crf_test$doc_id)
#' head(scores)
#' crf_test$label <- scores$label
#'
#' ## cleanup for CRAN
#' if(file.exists(model$file_model)) file.remove(model$file_model)
#' if(file.exists("modeldetails.txt")) file.remove("modeldetails.txt")
#' }
#' \donttest{
#' ##
#' ## More detailed example where text data was annotated with the webapp in the package
#' ## This data is joined with a tokenised dataset to construct the training data which
#' ## is further enriched with attributes of upos/lemma in the neighbourhood
#' ##
#' \dontshow{if(require(udpipe))\{}
#' library(udpipe)
#' data(airbnb_chunks, package = "crfsuite")
#' udmodel <- udpipe_download_model("dutch-lassysmall")
#' if(!udmodel$download_failed){
#' udmodel <- udpipe_load_model(udmodel$file_model)
#' airbnb_tokens <- udpipe(x = unique(airbnb_chunks[, c("doc_id", "text")]),
#' object = udmodel)
#' x <- merge(airbnb_chunks, airbnb_tokens)
#' x <- crf_cbind_attributes(x, terms = c("upos", "lemma"), by = "doc_id")
#' model <- crf(y = x$chunk_entity,
#' x = x[, grep("upos|lemma", colnames(x), value = TRUE)],
#' group = x$doc_id,
#' method = "lbfgs", options = list(max_iterations = 5))
#' stats <- summary(model)
#' stats
#' plot(stats$iterations$loss, type = "b", xlab = "Iteration", ylab = "Loss")
#' scores <- predict(model,
#' newdata = x[, grep("upos|lemma", colnames(x))],
#' group = x$doc_id)
#' head(scores)
#' }
#' \dontshow{\} # End of main if statement running only if the required packages are installed}
#' }
crf <- function(x, y, group,
method = c("lbfgs", "l2sgd", "averaged-perceptron", "passive-aggressive", "arow"),
options = crf_options(method)$default,
embeddings,
file = "annotator.crfsuite", trace = FALSE, FUN = identity, ...){
type <- "crf1d"
file <- normalizePath(path.expand(file), mustWork = FALSE)
trace <- as.integer(trace)
method <- match.arg(method)
x <- as.matrix(x)
stopifnot(length(group) == length(y))
stopifnot(nrow(x) == length(y))
if(!is.integer(group)){
group <- as.integer(factor(group))
}
options <- lapply(options, as.character)
logfile <- tempfile(pattern = sprintf("CRFSUITE_TRAINER_LOG_RUN_%s_", tools::file_path_sans_ext(basename(file))),
tmpdir = getwd(), fileext = ".log")
Sys.setenv("CRFSUITE_TRAINER_LOG" = logfile)
f <- Sys.getenv("CRFSUITE_TRAINER_LOG")
if(trace){
cat(sprintf("CRFsuite training progress logged to file %s", f), sep = "\n")
}
on.exit(file.remove(f))
if(missing(embeddings)){
embeddings <- matrix(numeric(), nrow = 0, ncol = 0)
}else{
embeddings <- as.matrix(embeddings)
stopifnot(nrow(embeddings) == nrow(x))
if(is.null(colnames(embeddings))){
colnames(embeddings) <- sprintf("embedding_dim%s", seq_len(ncol(embeddings)))
}
}
model <- crfsuite_model_build(file_model = file,
doc_id = group, y = y, x = x, embeddings = embeddings,
options = options, method = method, type = type,
trace = trace)
model$attribute_names <- colnames(x)
model$log <- readLines(f)
model$FUN <- FUN
model$ldots <- list(...)
class(model) <- "crf"
if(trace){
cat("CRFsuite training progress log file removed", sep = "\n")
}
model
}
#' @title Convert a model built with CRFsuite to an object of class crf
#' @description If you have a model built with CRFsuite either by this R package
#' or by another software library which wraps CRFsuite (e.g. Python/Java), you can
#' convert it to an object of class \code{crf} which this package can use
#' to inspect the model and to use it for prediction (if you can mimic the way the attributes are created).\cr
#' This is for expert use only.
#' @param file the path to a file on disk containing the CRFsuite model
#' @param ... other arguments which can be set except the path to the file, namely method, type, options, attribute_names, log (expert use only)
#' @return an object of class \code{crf}
#' @export
as.crf <- function(file, ...){
ldots <- list(...)
stopifnot(file.exists(file))
object <- crfsuite_model(file)
for(element in c("method", "type", "options", "attribute_names", "log", "FUN", "ldots")){
object[[element]] <- ldots[[element]]
}
class(object) <- "crf"
object
}
#' @export
print.crf <- function(x, ...){
fsize <- file.info(x$file_model)$size
cat(sprintf("Conditional Random Field saved at %s", x$file_model), sep = "\n")
cat(sprintf(" size of the model in Mb: %s", round(fsize / (2^20), 2)), sep = "\n")
cat(sprintf(" number of categories: %s", length(x$labels)), sep = "\n")
cat(sprintf(" category labels: %s", paste(x$labels, collapse = ", ")), sep = "\n")
}
#' @export
summary.crf <- function(object, file, ...){
stopifnot(file.exists(object$file_model))
out <- list()
out$active <- list()
out$active$features <- as.numeric(gsub("(^Number of active features: )(.+) (.+)$", "\\2", grep("^Number of active features: ", object$log, value = TRUE)))
out$active$attributes <- as.numeric(gsub("(^Number of active attributes: )(.+) (.+)$", "\\2", grep("^Number of active attributes: ", object$log, value = TRUE)))
out$active$labels <- as.numeric(gsub("(^Number of active labels: )(.+) (.+)$", "\\2", grep("^Number of active labels: ", object$log, value = TRUE)))
out$iterations <- list()
out$iterations$nr <- grep("^\\*\\*\\*\\*\\* (Iteration)|(Epoch) #", object$log, value = TRUE)
out$iterations$nr <- as.numeric(sapply(strsplit(out$iterations$nr, split = "#"), FUN=function(x) strsplit(tail(x, 1), " ")[[1]][1]))
out$iterations$loss <- as.numeric(gsub("Loss: ", "", grep("^Loss: ", object$log, value=TRUE)))
out$iterations$feature_norm <- as.numeric(gsub("Feature norm: ", "", grep("^Feature norm: ", object$log, value=TRUE)))
out$iterations$error_norm <- as.numeric(gsub("Error norm: ", "", grep("^Error norm: ", object$log, value=TRUE)))
out$iterations$active_features <- as.numeric(gsub("Active features: ", "", grep("^Active features: ", object$log, value=TRUE)))
out$iterations$line_search_trials <- as.numeric(gsub("Line search trials: ", "", grep("^Line search trials: ", object$log, value=TRUE)))
out$iterations$line_search_step <- as.numeric(gsub("Line search step: ", "", grep("^Line search step: ", object$log, value=TRUE)))
out$iterations$feature_l2norm <- as.numeric(gsub("Feature L2-norm: ", "", grep("^Feature L2-norm: ", object$log, value=TRUE)))
out$iterations$improvement_ratio <- as.numeric(gsub("Improvement ratio: ", "", grep("^Improvement ratio: ", object$log, value=TRUE)))
out$iterations$learning_rate_eta <- as.numeric(gsub("Learning rate .+: ", "", grep("^Learning rate .+: ", object$log, value=TRUE)))
out$iterations$feature_updates <- as.numeric(gsub("Total number of feature updates: ", "", grep("^Total number of feature updates: ", object$log, value=TRUE)))
out$iterations$seconds <- as.numeric(gsub("Seconds required for this iteration: ", "", grep("^Seconds required for this iteration: ", object$log, value=TRUE)))
out$iterations <- out$iterations[sapply(out$iterations, FUN=function(x) length(x) > 0)]
out$last_iteration <- tail(grep("^\\*\\*\\*\\*\\* (Iteration)|(Epoch) #", object$log), 1)
if(length(out$last_iteration) > 0){
out$last_iteration <- tail(object$log, length(object$log) - out$last_iteration)
out$last_iteration <- out$last_iteration[1:(head(which(out$last_iteration == ""), 1)-1)]
cat("Summary statistics of last iteration: ", sep = "\n")
cat(out$last_iteration, sep = "\n")
}
tempfile <- tempfile(pattern = "crfsuite_", fileext = ".txt")
cat(sprintf("\nDumping summary of the model to file %s", tempfile), sep = "\n")
crfsuite_model_dump(object$file_model, tempfile)
if(!missing(file)){
cat(sprintf("Copying summary of the model to file %s in folder %s", file, ifelse(basename(file) == file, getwd(), dirname(file))), sep = "\n")
file.copy(from = tempfile, to = file, overwrite = TRUE)
}
invisible(out)
}
#' @title Predict the label sequence based on the Conditional Random Field
#' @description Predict the label sequence based on the Conditional Random Field
#' @param object an object of class crf as returned by \code{\link{crf}}
#' @param newdata a character matrix of data containing attributes about the label sequence \code{y} or an object which can be coerced to a character matrix.
#' This data should be provided in the same format as was used for training the model
#' @param embeddings a matrix with the same number of rows as \code{x} and in the same order with numeric information used to predict
#' @param group an integer or character vector of the same length as nrow \code{newdata} indicating the group the sequence \code{y} belongs to (e.g. a document or sentence identifier)
#' @param type either 'marginal' or 'sequence' to get predictions at the level of \code{newdata} or a the level of the sequence \code{group}. Defaults to \code{'marginal'}
#' @param trace a logical indicating to show the trace of the labelling output. Defaults to \code{FALSE}.
#' @param ... not used
#' @export
#' @return
#' If \code{type} is 'marginal': a data.frame with columns label and marginal containing the viterbi decoded predicted label and marginal probability. \cr
#' If \code{type} is 'sequence': a data.frame with columns group and probability containing for each sequence group the probability of the sequence.
#' @seealso \code{\link{crf}}
#' @examples
#' \donttest{
#' \dontshow{if(require(udpipe))\{}
#' library(udpipe)
#' data(airbnb_chunks, package = "crfsuite")
#' udmodel <- udpipe_download_model("dutch-lassysmall")
#' udmodel <- udpipe_load_model(udmodel$file_model)
#' airbnb_tokens <- unique(airbnb_chunks[, c("doc_id", "text")])
#' airbnb_tokens <- udpipe_annotate(udmodel,
#' x = airbnb_tokens$text,
#' doc_id = airbnb_tokens$doc_id)
#' airbnb_tokens <- as.data.frame(airbnb_tokens)
#' x <- merge(airbnb_chunks, airbnb_tokens)
#' x <- crf_cbind_attributes(x, terms = c("upos", "lemma"), by = "doc_id")
#' model <- crf(y = x$chunk_entity,
#' x = x[, grep("upos|lemma", colnames(x))],
#' group = x$doc_id,
#' method = "lbfgs", options = list(max_iterations = 5))
#' scores <- predict(model,
#' newdata = x[, grep("upos|lemma", colnames(x))],
#' group = x$doc_id, type = "marginal")
#' head(scores)
#' scores <- predict(model,
#' newdata = x[, grep("upos|lemma", colnames(x))],
#' group = x$doc_id, type = "sequence")
#' head(scores)
#'
#'
#' ## cleanup for CRAN
#' file.remove(model$file_model)
#' file.remove("modeldetails.txt")
#' file.remove(udmodel$file)
#' \dontshow{\} # End of main if statement running only if the required packages are installed}
#' }
predict.crf <- function(object, newdata, embeddings, group, type = c("marginal", "sequence"), trace = FALSE, ...){
stopifnot(file.exists(object$file_model))
trace <- as.integer(trace)
newdata <- as.matrix(newdata)
stopifnot(nrow(newdata) == length(group))
type <- match.arg(type)
if(!is.integer(group)){
recodegroup <- TRUE
group <- factor(group)
levs <- levels(group)
group <- as.integer(group)
}else{
recodegroup <- FALSE
}
if(missing(embeddings)){
embeddings <- matrix(numeric(), nrow = 0, ncol = 0)
}else{
embeddings <- as.matrix(embeddings)
stopifnot(nrow(embeddings) == nrow(newdata))
if(is.null(colnames(embeddings))){
colnames(embeddings) <- sprintf("embedding_dim%s", seq_len(ncol(embeddings)))
}
}
scores <- crfsuite_predict(file_model = object$file_model, doc_id = group, x = newdata, embeddings = embeddings, trace = trace)
if(type == "marginal"){
scores <- scores$viterbi
}else if(type == "sequence"){
scores <- scores$sequence
if(recodegroup){
scores$group <- levs[match(scores$group, table = seq_along(levs))]
}
}
scores
}
#' @title Conditional Random Fields parameters
#' @description Conditional Random Fields parameters
#' @param method character string with the type of training method. Either one of:
#' \itemize{
#' \item{lbfgs: }{L-BFGS with L1/L2 regularization}
#' \item{l2sgd: }{SGD with L2-regularization}
#' \item{averaged-perceptron: }{Averaged Perceptron}
#' \item{passive-aggressive: }{Passive Aggressive}
#' \item{arow: }{Adaptive Regularization of Weights (AROW)}
#' }
#' @return a list with elements
#' \itemize{
#' \item{method: }{The training method}
#' \item{type: }{The type of graphical model which is always set crf1d: Linear-chain (first-order Markov) CRF}
#' \item{params: }{A data.frame with fields arg, arg_default and description indicating the possible hyperparameters of the algorithm, the default values and the description}
#' \item{default: }{A list of default values which can be used to pass on to the \code{options} argument of \code{\link{crf}}}
#' }
#' @export
#' @examples
#' # L-BFGS with L1/L2 regularization
#' opts <- crf_options("lbfgs")
#' str(opts)
#'
#' # SGD with L2-regularization
#' crf_options("l2sgd")
#'
#' # Averaged Perceptron
#' crf_options("averaged-perceptron")
#'
#' # Passive Aggressive
#' crf_options("passive-aggressive")
#'
#' # Adaptive Regularization of Weights (AROW)
#' crf_options("arow")
crf_options <- function(method = c("lbfgs", "l2sgd", "averaged-perceptron", "passive-aggressive", "arow")){
type <- "crf1d"
method <- match.arg(method)
params <- crfsuite_model_parameters(method, type)
params$default <- structure(as.list(params$params$arg_default),
.Names = params$params$arg)
params$default <- lapply(params$default, FUN=function(x){
suppressWarnings({
value <- as.numeric(x)
value <- ifelse(is.na(value), x, value)
})
value
})
params
}
#' @title Functionality allowing to tune a crfsuite model using caret
#' @description The object \code{crf_caretmethod} contains functionality to tune a crf model using caret.
#' Each list elment of \code{crf_caretmethod} is a list of functions
#' which can be passed on to the \code{method} argument of \code{caret::train} to tune the hyperparameters of the crfsuite model.
#' @export
#' @rdname crf_caretmethod
#' @format see details
#' @details If you want to tune the hyperparameters of a crfsuite model
#' (see \code{\link{crf_options}} and the \code{options} argument of \code{\link{crf}}), you can use the \code{caret} package. \cr
#' In order to facilitate this tuning, an object called \code{crf_caretmethod} has been made available.
#' The object \code{crf_caretmethod} is a list with 6 elements, where each of these 6 elements can be used in
#' tuning the CRF hyperparemeters by passing it on to the \code{method} argument of the \code{train} function of the \code{caret} package.\cr
#' The list has elements 'default', 'lbfgs', 'l2sgd', 'averaged_perceptron', 'passive_aggressive' and 'arow'.
#' Each list element corresponds to arguments that you need to tune for each \code{method} as used in \code{\link{crf}}. \cr
#' For \code{crf_caretmethod}
#' \enumerate{
#' \item lbfgs: Tuning across all hyperparameters for method lbfgs: L-BFGS with L1/L2 regularization
#' \item l2sgd: Tuning across all hyperparameters for method l2sgd: SGD with L2-regularization
#' \item averaged_perceptron: Tuning across all hyperparameters for method averaged-perceptron: Averaged Perceptron
#' \item passive_aggressive: Tuning across all hyperparameters for method passive-aggressive: Passive Aggressive
#' \item arow: Tuning across all hyperparameters for method arow: Adaptive Regularization of Weights (AROW)
#' \item default: Tune over the hyperparameters feature.minfreq, feature.possible_states, feature.possible_transitions, max_iterations. While tuning these, it uses the default hyperparameters for each method. This tuning allows you to compare the 5 methods.
#' }
#' For details on the hyperparameter definitions: see \code{\link{crf_options}}
crf_caretmethod <- list(
default = list(library = "crfsuite", type = "Classification",
parameters = data.frame(parameter = c("method", "feature.minfreq", "feature.possible_states", "feature.possible_transitions", "max_iterations"),
class = c("character", "integer", "logical", "logical", "integer"),
label = c("Training method", "The minimum frequency of features.", "Generate all possible state features", "Generate all possible transition features", "Maximum number of iterations"),
stringsAsFactors = FALSE),
grid = function(x, y, len = NULL, search = "grid") .NotYetImplemented(),
fit = function(x, y, wts, param, lev, last, weights, classProbs, ...){
opts <- crf_options(method = param$method)$default
if(length(intersect("method", names(param))) == 0) stop("method should always be one of the arguments")
extraparams <- setdiff(names(param), "method")
extraparams <- intersect(extraparams, names(opts))
opts[extraparams] <- param[extraparams]
cat(sprintf("%s - building model for setting: %s, %s", Sys.time(), param$method, paste(mapply(extraparams, param[extraparams], FUN=function(key, value) paste(key, value, sep = "=")), collapse = ", ")), sep = "\n")
crf(x = x[, -1], y = as.character(y), group = x[, 1], method = param$method, options = opts, ...)
},
predict = function(modelFit, newdata, submodels = NULL){
predict(modelFit, newdata = newdata[, -1], group = newdata[, 1], type = "marginal")$label
},
prob = NULL)
)
crf_caretmethod$lbfgs <- crf_caretmethod$default
crf_caretmethod$l2sgd <- crf_caretmethod$default
crf_caretmethod$averaged_perceptron <- crf_caretmethod$default
crf_caretmethod$passive_aggressive <- crf_caretmethod$default
crf_caretmethod$arow <- crf_caretmethod$default
crf_caretmethod$lbfgs$parameters <- data.frame(parameter = c("method",
"feature.minfreq", "feature.possible_states", "feature.possible_transitions",
"c1", "c2", "max_iterations", "num_memories", "epsilon", "period",
"delta", "linesearch", "max_linesearch"),
class = c("character", "integer", "logical", "logical", "numeric", "numeric", "integer", "integer", "numeric", "integer", "numeric", "character", "integer"),
label = c("Training method",
"The minimum frequency of features.", "Force to generate possible state features.",
"Force to generate possible transition features.", "Coefficient for L1 regularization.",
"Coefficient for L2 regularization.", "The maximum number of iterations for L-BFGS optimization.",
"The number of limited memories for approximating the inverse hessian matrix.",
"Epsilon for testing the convergence of the objective.", "The duration of iterations to test the stopping criterion.",
"The threshold for the stopping criterion; an L-BFGS iteration stops when the\nimprovement of the log likelihood over the last ${period} iterations is no\ngreater than this threshold.",
"The line search algorithm used in L-BFGS updates:\n{ 'MoreThuente': More and Thuente's method,\n 'Backtracking': Backtracking method with regular Wolfe condition,\n 'StrongBacktracking': Backtracking method with strong Wolfe condition\n}\n",
"The maximum number of trials for the line search algorithm."),
stringsAsFactors = FALSE)
crf_caretmethod$l2sgd$parameters <- data.frame(parameter = c("method",
"feature.minfreq", "feature.possible_states", "feature.possible_transitions",
"c2", "max_iterations", "period", "delta", "calibration.eta",
"calibration.rate", "calibration.samples", "calibration.candidates",
"calibration.max_trials"),
class = c("character", "integer", "logical", "logical", "numeric", "integer", "integer", "numeric", "numeric", "numeric", "integer", "integer", "integer"),
label = c("Training method",
"The minimum frequency of features.", "Force to generate possible state features.",
"Force to generate possible transition features.", "Coefficient for L2 regularization.",
"The maximum number of iterations (epochs) for SGD optimization.",
"The duration of iterations to test the stopping criterion.",
"The threshold for the stopping criterion; an optimization process stops when\nthe improvement of the log likelihood over the last ${period} iterations is no\ngreater than this threshold.",
"The initial value of learning rate (eta) used for calibration.",
"The rate of increase/decrease of learning rate for calibration.",
"The number of instances used for calibration.", "The number of candidates of learning rate.",
"The maximum number of trials of learning rates for calibration."),
stringsAsFactors = FALSE)
crf_caretmethod$averaged_perceptron$parameters <- data.frame(parameter = c("method",
"feature.minfreq", "feature.possible_states", "feature.possible_transitions",
"max_iterations", "epsilon"),
class = c("character", "integer", "logical", "logical", "integer", "numeric"),
label = c("Training method",
"The minimum frequency of features.", "Force to generate possible state features.",
"Force to generate possible transition features.", "The maximum number of iterations.",
"The stopping criterion (the ratio of incorrect label predictions)."),
stringsAsFactors = FALSE)
crf_caretmethod$passive_aggressive$parameters <- data.frame(parameter = c("method",
"feature.minfreq", "feature.possible_states", "feature.possible_transitions",
"type", "c", "error_sensitive", "averaging", "max_iterations",
"epsilon"),
class = c("character", "integer", "logical", "logical", "integer", "logical", "logical", "logical", "integer", "numeric"),
label = c("Training method",
"The minimum frequency of features.", "Force to generate possible state features.",
"Force to generate possible transition features.", "The strategy for updating feature weights: {\n 0: PA without slack variables,\n 1: PA type I,\n 2: PA type II\n}.\n",
"The aggressiveness parameter.", "Consider the number of incorrect labels to the cost function.",
"Compute the average of feature weights (similarly to Averaged Perceptron).",
"The maximum number of iterations.", "The stopping criterion (the mean loss)."),
stringsAsFactors = FALSE)
crf_caretmethod$arow$parameters <- data.frame(parameter = c("method",
"feature.minfreq", "feature.possible_states", "feature.possible_transitions",
"variance", "gamma", "max_iterations", "epsilon"),
class = c("character", "integer", "logical", "logical", "numeric", "numeric", "integer", "numeric"),
label = c("Training method",
"The minimum frequency of features.", "Force to generate possible state features.",
"Force to generate possible transition features.", "The initial variance of every feature weight.",
"Tradeoff parameter.", "The maximum number of iterations.", "The stopping criterion (the mean loss)."),
stringsAsFactors = FALSE)
Try the crfsuite package in your browser
Any scripts or data that you put into this service are public.
crfsuite documentation built on Sept. 17, 2023, 1:06 a.m.
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
Defines functions crf_options predict.crf summary.crf print.crf as.crf crf
Documented in as.crf crf crf_options predict.crf
#' @title Linear-chain Conditional Random Field
#' @description Fits a Linear-chain (first-order Markov) CRF on the provided label sequence and saves it on disk in order to do sequence labelling.
#' @param x a character matrix of data containing attributes about the label sequence \code{y} or an object which can be coerced to a character matrix.
#' It is important to note that an attribute which has the same value in a different column is considered the same.
#' @param y a character vector with the sequence of labels to model
#' @param group an integer or character vector of the same length as \code{y} indicating the group the sequence \code{y} belongs to (e.g. a document or sentence identifier)
#' @param method character string with the type of training method. Either one of:
#' \itemize{
#' \item{lbfgs: }{L-BFGS with L1/L2 regularization}
#' \item{l2sgd: }{SGD with L2-regularization}
#' \item{averaged-perceptron: }{Averaged Perceptron}
#' \item{passive-aggressive: }{Passive Aggressive}
#' \item{arow: }{Adaptive Regularization of Weights (AROW)}
#' }
#' @param options a list of options to provide to the training algorithm. See \code{\link{crf_options}} for possible options and the example below on how to provide them.
#' @param embeddings a matrix with the same number of rows as \code{x} and in the same order with numeric information used in model building (experimental)
#' @param file a character string with the path to the file on disk where the CRF model will be stored.
#' @param trace a logical indicating to show the trace of the training output. Defaults to \code{FALSE}.
#' @param FUN a function which can be applied on raw text in order to obtain the attribute matrix used in \code{predict.crf}. Currently not used yet.
#' @param ... arguments to FUN. Currently not used yet.
#' @return an object of class crf which is a list with elements
#' \itemize{
#' \item{method: }{The training method}
#' \item{type: }{The type of graphical model which is always set crf1d: Linear-chain (first-order Markov) CRF}
#' \item{labels: }{The training labels}
#' \item{options: }{A data.frame with the training options provided to the algorithm}
#' \item{file_model: }{The path where the CRF model is stored}
#' \item{attribute_names: }{The column names of \code{x}}
#' \item{log: }{The training log of the algorithm}
#' \item{FUN: }{The argument passed on to FUN}
#' \item{ldots: }{A list with the arguments passed on to ...}
#' }
#' @references More details about this model is available at \url{http://www.chokkan.org/software/crfsuite/}.
#' @export
#' @seealso \code{\link{predict.crf}}
#' @examples
#' ## Download modeldata (conll 2002 shared task in Dutch)
#' \donttest{
#' x <- ner_download_modeldata("conll2002-nl")
#' }
#' # for CRAN only - word on a subset of the data
#' x <- ner_download_modeldata("conll2002-nl", docs = 10)
#' if(is.data.frame(x)){
#' ##
#' ## Build Named Entity Recognition model on conll2002-nl
#' ##
#' x$pos <- txt_sprintf("Parts of Speech: %s", x$pos)
#' x$token <- txt_sprintf("Token: %s", x$token)
#' crf_train <- subset(x, data == "ned.train")
#' crf_test <- subset(x, data == "testa")
#'
#' model <- crf(y = crf_train$label,
#' x = crf_train[, c("token", "pos")],
#' group = crf_train$doc_id,
#' method = "lbfgs",
#' options = list(max_iterations = 3, feature.minfreq = 5,
#' c1 = 0, c2 = 1))
#' model
#' weights <- coefficients(model)
#' head(weights$states, n = 20)
#' head(weights$transitions, n = 20)
#' stats <- summary(model, "modeldetails.txt")
#' stats
#' plot(stats$iterations$loss)
#'
#' ## Use the CRF model to label a sequence
#' scores <- predict(model,
#' newdata = crf_test[, c("token", "pos")],
#' group = crf_test$doc_id)
#' head(scores)
#' crf_test$label <- scores$label
#'
#' ## cleanup for CRAN
#' if(file.exists(model$file_model)) file.remove(model$file_model)
#' if(file.exists("modeldetails.txt")) file.remove("modeldetails.txt")
#' }
#' \donttest{
#' ##
#' ## More detailed example where text data was annotated with the webapp in the package
#' ## This data is joined with a tokenised dataset to construct the training data which
#' ## is further enriched with attributes of upos/lemma in the neighbourhood
#' ##
#' \dontshow{if(require(udpipe))\{}
#' library(udpipe)
#' data(airbnb_chunks, package = "crfsuite")
#' udmodel <- udpipe_download_model("dutch-lassysmall")
#' if(!udmodel$download_failed){
#' udmodel <- udpipe_load_model(udmodel$file_model)
#' airbnb_tokens <- udpipe(x = unique(airbnb_chunks[, c("doc_id", "text")]),
#' object = udmodel)
#' x <- merge(airbnb_chunks, airbnb_tokens)
#' x <- crf_cbind_attributes(x, terms = c("upos", "lemma"), by = "doc_id")
#' model <- crf(y = x$chunk_entity,
#' x = x[, grep("upos|lemma", colnames(x), value = TRUE)],
#' group = x$doc_id,
#' method = "lbfgs", options = list(max_iterations = 5))
#' stats <- summary(model)
#' stats
#' plot(stats$iterations$loss, type = "b", xlab = "Iteration", ylab = "Loss")
#' scores <- predict(model,
#' newdata = x[, grep("upos|lemma", colnames(x))],
#' group = x$doc_id)
#' head(scores)
#' }
#' \dontshow{\} # End of main if statement running only if the required packages are installed}
#' }
crf <- function(x, y, group,
method = c("lbfgs", "l2sgd", "averaged-perceptron", "passive-aggressive", "arow"),
options = crf_options(method)$default,
embeddings,
file = "annotator.crfsuite", trace = FALSE, FUN = identity, ...){
type <- "crf1d"
file <- normalizePath(path.expand(file), mustWork = FALSE)
trace <- as.integer(trace)
method <- match.arg(method)
x <- as.matrix(x)
stopifnot(length(group) == length(y))
stopifnot(nrow(x) == length(y))
if(!is.integer(group)){
group <- as.integer(factor(group))
}
options <- lapply(options, as.character)
logfile <- tempfile(pattern = sprintf("CRFSUITE_TRAINER_LOG_RUN_%s_", tools::file_path_sans_ext(basename(file))),
tmpdir = getwd(), fileext = ".log")
Sys.setenv("CRFSUITE_TRAINER_LOG" = logfile)
f <- Sys.getenv("CRFSUITE_TRAINER_LOG")
if(trace){
cat(sprintf("CRFsuite training progress logged to file %s", f), sep = "\n")
}
on.exit(file.remove(f))
if(missing(embeddings)){
embeddings <- matrix(numeric(), nrow = 0, ncol = 0)
}else{
embeddings <- as.matrix(embeddings)
stopifnot(nrow(embeddings) == nrow(x))
if(is.null(colnames(embeddings))){
colnames(embeddings) <- sprintf("embedding_dim%s", seq_len(ncol(embeddings)))
}
}
model <- crfsuite_model_build(file_model = file,
doc_id = group, y = y, x = x, embeddings = embeddings,
options = options, method = method, type = type,
trace = trace)
model$attribute_names <- colnames(x)
model$log <- readLines(f)
model$FUN <- FUN
model$ldots <- list(...)
class(model) <- "crf"
if(trace){
cat("CRFsuite training progress log file removed", sep = "\n")
}
model
}
#' @title Convert a model built with CRFsuite to an object of class crf
#' @description If you have a model built with CRFsuite either by this R package
#' or by another software library which wraps CRFsuite (e.g. Python/Java), you can
#' convert it to an object of class \code{crf} which this package can use
#' to inspect the model and to use it for prediction (if you can mimic the way the attributes are created).\cr
#' This is for expert use only.
#' @param file the path to a file on disk containing the CRFsuite model
#' @param ... other arguments which can be set except the path to the file, namely method, type, options, attribute_names, log (expert use only)
#' @return an object of class \code{crf}
#' @export
as.crf <- function(file, ...){
ldots <- list(...)
stopifnot(file.exists(file))
object <- crfsuite_model(file)
for(element in c("method", "type", "options", "attribute_names", "log", "FUN", "ldots")){
object[[element]] <- ldots[[element]]
}
class(object) <- "crf"
object
}
#' @export
print.crf <- function(x, ...){
fsize <- file.info(x$file_model)$size
cat(sprintf("Conditional Random Field saved at %s", x$file_model), sep = "\n")
cat(sprintf(" size of the model in Mb: %s", round(fsize / (2^20), 2)), sep = "\n")
cat(sprintf(" number of categories: %s", length(x$labels)), sep = "\n")
cat(sprintf(" category labels: %s", paste(x$labels, collapse = ", ")), sep = "\n")
}
#' @export
summary.crf <- function(object, file, ...){
stopifnot(file.exists(object$file_model))
out <- list()
out$active <- list()
out$active$features <- as.numeric(gsub("(^Number of active features: )(.+) (.+)$", "\\2", grep("^Number of active features: ", object$log, value = TRUE)))
out$active$attributes <- as.numeric(gsub("(^Number of active attributes: )(.+) (.+)$", "\\2", grep("^Number of active attributes: ", object$log, value = TRUE)))
out$active$labels <- as.numeric(gsub("(^Number of active labels: )(.+) (.+)$", "\\2", grep("^Number of active labels: ", object$log, value = TRUE)))
out$iterations <- list()
out$iterations$nr <- grep("^\\*\\*\\*\\*\\* (Iteration)|(Epoch) #", object$log, value = TRUE)
out$iterations$nr <- as.numeric(sapply(strsplit(out$iterations$nr, split = "#"), FUN=function(x) strsplit(tail(x, 1), " ")[[1]][1]))
out$iterations$loss <- as.numeric(gsub("Loss: ", "", grep("^Loss: ", object$log, value=TRUE)))
out$iterations$feature_norm <- as.numeric(gsub("Feature norm: ", "", grep("^Feature norm: ", object$log, value=TRUE)))
out$iterations$error_norm <- as.numeric(gsub("Error norm: ", "", grep("^Error norm: ", object$log, value=TRUE)))
out$iterations$active_features <- as.numeric(gsub("Active features: ", "", grep("^Active features: ", object$log, value=TRUE)))
out$iterations$line_search_trials <- as.numeric(gsub("Line search trials: ", "", grep("^Line search trials: ", object$log, value=TRUE)))
out$iterations$line_search_step <- as.numeric(gsub("Line search step: ", "", grep("^Line search step: ", object$log, value=TRUE)))
out$iterations$feature_l2norm <- as.numeric(gsub("Feature L2-norm: ", "", grep("^Feature L2-norm: ", object$log, value=TRUE)))
out$iterations$improvement_ratio <- as.numeric(gsub("Improvement ratio: ", "", grep("^Improvement ratio: ", object$log, value=TRUE)))
out$iterations$learning_rate_eta <- as.numeric(gsub("Learning rate .+: ", "", grep("^Learning rate .+: ", object$log, value=TRUE)))
out$iterations$feature_updates <- as.numeric(gsub("Total number of feature updates: ", "", grep("^Total number of feature updates: ", object$log, value=TRUE)))
out$iterations$seconds <- as.numeric(gsub("Seconds required for this iteration: ", "", grep("^Seconds required for this iteration: ", object$log, value=TRUE)))
out$iterations <- out$iterations[sapply(out$iterations, FUN=function(x) length(x) > 0)]
out$last_iteration <- tail(grep("^\\*\\*\\*\\*\\* (Iteration)|(Epoch) #", object$log), 1)
if(length(out$last_iteration) > 0){
out$last_iteration <- tail(object$log, length(object$log) - out$last_iteration)
out$last_iteration <- out$last_iteration[1:(head(which(out$last_iteration == ""), 1)-1)]
cat("Summary statistics of last iteration: ", sep = "\n")
cat(out$last_iteration, sep = "\n")
}
tempfile <- tempfile(pattern = "crfsuite_", fileext = ".txt")
cat(sprintf("\nDumping summary of the model to file %s", tempfile), sep = "\n")
crfsuite_model_dump(object$file_model, tempfile)
if(!missing(file)){
cat(sprintf("Copying summary of the model to file %s in folder %s", file, ifelse(basename(file) == file, getwd(), dirname(file))), sep = "\n")
file.copy(from = tempfile, to = file, overwrite = TRUE)
}
invisible(out)
}
#' @title Predict the label sequence based on the Conditional Random Field
#' @description Predict the label sequence based on the Conditional Random Field
#' @param object an object of class crf as returned by \code{\link{crf}}
#' @param newdata a character matrix of data containing attributes about the label sequence \code{y} or an object which can be coerced to a character matrix.
#' This data should be provided in the same format as was used for training the model
#' @param embeddings a matrix with the same number of rows as \code{x} and in the same order with numeric information used to predict
#' @param group an integer or character vector of the same length as nrow \code{newdata} indicating the group the sequence \code{y} belongs to (e.g. a document or sentence identifier)
#' @param type either 'marginal' or 'sequence' to get predictions at the level of \code{newdata} or a the level of the sequence \code{group}. Defaults to \code{'marginal'}
#' @param trace a logical indicating to show the trace of the labelling output. Defaults to \code{FALSE}.
#' @param ... not used
#' @export
#' @return
#' If \code{type} is 'marginal': a data.frame with columns label and marginal containing the viterbi decoded predicted label and marginal probability. \cr
#' If \code{type} is 'sequence': a data.frame with columns group and probability containing for each sequence group the probability of the sequence.
#' @seealso \code{\link{crf}}
#' @examples
#' \donttest{
#' \dontshow{if(require(udpipe))\{}
#' library(udpipe)
#' data(airbnb_chunks, package = "crfsuite")
#' udmodel <- udpipe_download_model("dutch-lassysmall")
#' udmodel <- udpipe_load_model(udmodel$file_model)
#' airbnb_tokens <- unique(airbnb_chunks[, c("doc_id", "text")])
#' airbnb_tokens <- udpipe_annotate(udmodel,
#' x = airbnb_tokens$text,
#' doc_id = airbnb_tokens$doc_id)
#' airbnb_tokens <- as.data.frame(airbnb_tokens)
#' x <- merge(airbnb_chunks, airbnb_tokens)
#' x <- crf_cbind_attributes(x, terms = c("upos", "lemma"), by = "doc_id")
#' model <- crf(y = x$chunk_entity,
#' x = x[, grep("upos|lemma", colnames(x))],
#' group = x$doc_id,
#' method = "lbfgs", options = list(max_iterations = 5))
#' scores <- predict(model,
#' newdata = x[, grep("upos|lemma", colnames(x))],
#' group = x$doc_id, type = "marginal")
#' head(scores)
#' scores <- predict(model,
#' newdata = x[, grep("upos|lemma", colnames(x))],
#' group = x$doc_id, type = "sequence")
#' head(scores)
#'
#'
#' ## cleanup for CRAN
#' file.remove(model$file_model)
#' file.remove("modeldetails.txt")
#' file.remove(udmodel$file)
#' \dontshow{\} # End of main if statement running only if the required packages are installed}
#' }
predict.crf <- function(object, newdata, embeddings, group, type = c("marginal", "sequence"), trace = FALSE, ...){
stopifnot(file.exists(object$file_model))
trace <- as.integer(trace)
newdata <- as.matrix(newdata)
stopifnot(nrow(newdata) == length(group))
type <- match.arg(type)
if(!is.integer(group)){
recodegroup <- TRUE
group <- factor(group)
levs <- levels(group)
group <- as.integer(group)
}else{
recodegroup <- FALSE
}
if(missing(embeddings)){
embeddings <- matrix(numeric(), nrow = 0, ncol = 0)
}else{
embeddings <- as.matrix(embeddings)
stopifnot(nrow(embeddings) == nrow(newdata))
if(is.null(colnames(embeddings))){
colnames(embeddings) <- sprintf("embedding_dim%s", seq_len(ncol(embeddings)))
}
}
scores <- crfsuite_predict(file_model = object$file_model, doc_id = group, x = newdata, embeddings = embeddings, trace = trace)
if(type == "marginal"){
scores <- scores$viterbi
}else if(type == "sequence"){
scores <- scores$sequence
if(recodegroup){
scores$group <- levs[match(scores$group, table = seq_along(levs))]
}
}
scores
}
#' @title Conditional Random Fields parameters
#' @description Conditional Random Fields parameters
#' @param method character string with the type of training method. Either one of:
#' \itemize{
#' \item{lbfgs: }{L-BFGS with L1/L2 regularization}
#' \item{l2sgd: }{SGD with L2-regularization}
#' \item{averaged-perceptron: }{Averaged Perceptron}
#' \item{passive-aggressive: }{Passive Aggressive}
#' \item{arow: }{Adaptive Regularization of Weights (AROW)}
#' }
#' @return a list with elements
#' \itemize{
#' \item{method: }{The training method}
#' \item{type: }{The type of graphical model which is always set crf1d: Linear-chain (first-order Markov) CRF}
#' \item{params: }{A data.frame with fields arg, arg_default and description indicating the possible hyperparameters of the algorithm, the default values and the description}
#' \item{default: }{A list of default values which can be used to pass on to the \code{options} argument of \code{\link{crf}}}
#' }
#' @export
#' @examples
#' # L-BFGS with L1/L2 regularization
#' opts <- crf_options("lbfgs")
#' str(opts)
#'
#' # SGD with L2-regularization
#' crf_options("l2sgd")
#'
#' # Averaged Perceptron
#' crf_options("averaged-perceptron")
#'
#' # Passive Aggressive
#' crf_options("passive-aggressive")
#'
#' # Adaptive Regularization of Weights (AROW)
#' crf_options("arow")
crf_options <- function(method = c("lbfgs", "l2sgd", "averaged-perceptron", "passive-aggressive", "arow")){
type <- "crf1d"
method <- match.arg(method)
params <- crfsuite_model_parameters(method, type)
params$default <- structure(as.list(params$params$arg_default),
.Names = params$params$arg)
params$default <- lapply(params$default, FUN=function(x){
suppressWarnings({
value <- as.numeric(x)
value <- ifelse(is.na(value), x, value)
})
value
})
params
}
#' @title Functionality allowing to tune a crfsuite model using caret
#' @description The object \code{crf_caretmethod} contains functionality to tune a crf model using caret.
#' Each list elment of \code{crf_caretmethod} is a list of functions
#' which can be passed on to the \code{method} argument of \code{caret::train} to tune the hyperparameters of the crfsuite model.
#' @export
#' @rdname crf_caretmethod
#' @format see details
#' @details If you want to tune the hyperparameters of a crfsuite model
#' (see \code{\link{crf_options}} and the \code{options} argument of \code{\link{crf}}), you can use the \code{caret} package. \cr
#' In order to facilitate this tuning, an object called \code{crf_caretmethod} has been made available.
#' The object \code{crf_caretmethod} is a list with 6 elements, where each of these 6 elements can be used in
#' tuning the CRF hyperparemeters by passing it on to the \code{method} argument of the \code{train} function of the \code{caret} package.\cr
#' The list has elements 'default', 'lbfgs', 'l2sgd', 'averaged_perceptron', 'passive_aggressive' and 'arow'.
#' Each list element corresponds to arguments that you need to tune for each \code{method} as used in \code{\link{crf}}. \cr
#' For \code{crf_caretmethod}
#' \enumerate{
#' \item lbfgs: Tuning across all hyperparameters for method lbfgs: L-BFGS with L1/L2 regularization
#' \item l2sgd: Tuning across all hyperparameters for method l2sgd: SGD with L2-regularization
#' \item averaged_perceptron: Tuning across all hyperparameters for method averaged-perceptron: Averaged Perceptron
#' \item passive_aggressive: Tuning across all hyperparameters for method passive-aggressive: Passive Aggressive
#' \item arow: Tuning across all hyperparameters for method arow: Adaptive Regularization of Weights (AROW)
#' \item default: Tune over the hyperparameters feature.minfreq, feature.possible_states, feature.possible_transitions, max_iterations. While tuning these, it uses the default hyperparameters for each method. This tuning allows you to compare the 5 methods.
#' }
#' For details on the hyperparameter definitions: see \code{\link{crf_options}}
crf_caretmethod <- list(
default = list(library = "crfsuite", type = "Classification",
parameters = data.frame(parameter = c("method", "feature.minfreq", "feature.possible_states", "feature.possible_transitions", "max_iterations"),
class = c("character", "integer", "logical", "logical", "integer"),
label = c("Training method", "The minimum frequency of features.", "Generate all possible state features", "Generate all possible transition features", "Maximum number of iterations"),
stringsAsFactors = FALSE),
grid = function(x, y, len = NULL, search = "grid") .NotYetImplemented(),
fit = function(x, y, wts, param, lev, last, weights, classProbs, ...){
opts <- crf_options(method = param$method)$default
if(length(intersect("method", names(param))) == 0) stop("method should always be one of the arguments")
extraparams <- setdiff(names(param), "method")
extraparams <- intersect(extraparams, names(opts))
opts[extraparams] <- param[extraparams]
cat(sprintf("%s - building model for setting: %s, %s", Sys.time(), param$method, paste(mapply(extraparams, param[extraparams], FUN=function(key, value) paste(key, value, sep = "=")), collapse = ", ")), sep = "\n")
crf(x = x[, -1], y = as.character(y), group = x[, 1], method = param$method, options = opts, ...)
},
predict = function(modelFit, newdata, submodels = NULL){
predict(modelFit, newdata = newdata[, -1], group = newdata[, 1], type = "marginal")$label
},
prob = NULL)
)
crf_caretmethod$lbfgs <- crf_caretmethod$default
crf_caretmethod$l2sgd <- crf_caretmethod$default
crf_caretmethod$averaged_perceptron <- crf_caretmethod$default
crf_caretmethod$passive_aggressive <- crf_caretmethod$default
crf_caretmethod$arow <- crf_caretmethod$default
crf_caretmethod$lbfgs$parameters <- data.frame(parameter = c("method",
"feature.minfreq", "feature.possible_states", "feature.possible_transitions",
"c1", "c2", "max_iterations", "num_memories", "epsilon", "period",
"delta", "linesearch", "max_linesearch"),
class = c("character", "integer", "logical", "logical", "numeric", "numeric", "integer", "integer", "numeric", "integer", "numeric", "character", "integer"),
label = c("Training method",
"The minimum frequency of features.", "Force to generate possible state features.",
"Force to generate possible transition features.", "Coefficient for L1 regularization.",
"Coefficient for L2 regularization.", "The maximum number of iterations for L-BFGS optimization.",
"The number of limited memories for approximating the inverse hessian matrix.",
"Epsilon for testing the convergence of the objective.", "The duration of iterations to test the stopping criterion.",
"The threshold for the stopping criterion; an L-BFGS iteration stops when the\nimprovement of the log likelihood over the last ${period} iterations is no\ngreater than this threshold.",
"The line search algorithm used in L-BFGS updates:\n{ 'MoreThuente': More and Thuente's method,\n 'Backtracking': Backtracking method with regular Wolfe condition,\n 'StrongBacktracking': Backtracking method with strong Wolfe condition\n}\n",
"The maximum number of trials for the line search algorithm."),
stringsAsFactors = FALSE)
crf_caretmethod$l2sgd$parameters <- data.frame(parameter = c("method",
"feature.minfreq", "feature.possible_states", "feature.possible_transitions",
"c2", "max_iterations", "period", "delta", "calibration.eta",
"calibration.rate", "calibration.samples", "calibration.candidates",
"calibration.max_trials"),
class = c("character", "integer", "logical", "logical", "numeric", "integer", "integer", "numeric", "numeric", "numeric", "integer", "integer", "integer"),
label = c("Training method",
"The minimum frequency of features.", "Force to generate possible state features.",
"Force to generate possible transition features.", "Coefficient for L2 regularization.",
"The maximum number of iterations (epochs) for SGD optimization.",
"The duration of iterations to test the stopping criterion.",
"The threshold for the stopping criterion; an optimization process stops when\nthe improvement of the log likelihood over the last ${period} iterations is no\ngreater than this threshold.",
"The initial value of learning rate (eta) used for calibration.",
"The rate of increase/decrease of learning rate for calibration.",
"The number of instances used for calibration.", "The number of candidates of learning rate.",
"The maximum number of trials of learning rates for calibration."),
stringsAsFactors = FALSE)
crf_caretmethod$averaged_perceptron$parameters <- data.frame(parameter = c("method",
"feature.minfreq", "feature.possible_states", "feature.possible_transitions",
"max_iterations", "epsilon"),
class = c("character", "integer", "logical", "logical", "integer", "numeric"),
label = c("Training method",
"The minimum frequency of features.", "Force to generate possible state features.",
"Force to generate possible transition features.", "The maximum number of iterations.",
"The stopping criterion (the ratio of incorrect label predictions)."),
stringsAsFactors = FALSE)
crf_caretmethod$passive_aggressive$parameters <- data.frame(parameter = c("method",
"feature.minfreq", "feature.possible_states", "feature.possible_transitions",
"type", "c", "error_sensitive", "averaging", "max_iterations",
"epsilon"),
class = c("character", "integer", "logical", "logical", "integer", "logical", "logical", "logical", "integer", "numeric"),
label = c("Training method",
"The minimum frequency of features.", "Force to generate possible state features.",
"Force to generate possible transition features.", "The strategy for updating feature weights: {\n 0: PA without slack variables,\n 1: PA type I,\n 2: PA type II\n}.\n",
"The aggressiveness parameter.", "Consider the number of incorrect labels to the cost function.",
"Compute the average of feature weights (similarly to Averaged Perceptron).",
"The maximum number of iterations.", "The stopping criterion (the mean loss)."),
stringsAsFactors = FALSE)
crf_caretmethod$arow$parameters <- data.frame(parameter = c("method",
"feature.minfreq", "feature.possible_states", "feature.possible_transitions",
"variance", "gamma", "max_iterations", "epsilon"),
class = c("character", "integer", "logical", "logical", "numeric", "numeric", "integer", "numeric"),
label = c("Training method",
"The minimum frequency of features.", "Force to generate possible state features.",
"Force to generate possible transition features.", "The initial variance of every feature weight.",
"Tradeoff parameter.", "The maximum number of iterations.", "The stopping criterion (the mean loss)."),
stringsAsFactors = FALSE)
Try the crfsuite package in your browser
Any scripts or data that you put into this service are public.
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
Embedding an R snippet on your website
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.