ndl (version 0.2.18)

ndl-package: ndl

Description

ndl

Naive discriminative learning implements classification models based on the Rescorla-Wagner equations and the equilibrium equations of the Rescorla-Wagner equations. This package provides three kinds of functionality: (1) discriminative learning based directly on the Rescorla-Wagner equations, (2) a function implementing the naive discriminative reader, and a model for silent (single-word) reading, and (3) a classifier based on the equilibrium equations. The functions and datasets for the naive discriminative reader model make it possible to replicate the simulation results for Experiment 1 of Baayen et al. (2011). The classifier is provided to allow for comparisons between machine learning (svm, TiMBL, glm, random forests, etc.) and discrimination learning. Compared to standard classification algorithms, naive discriminative learning may overfit the data, albeit gracefully.

Details

The DESCRIPTION file: ndl ndl

For more detailed information on the core Rescorla-Wagner equations, see the functions RescorlaWagner and plot.RescorlaWagner, as well as the data sets danks, numbers (data courtesy of Michael Ramscar), and lexample (an example discussed in Baayen et al. 2011).

The functions for the naive discriminative learning (at the user level) are estimateWeights and estimateActivations. The relevant data sets are serbian, serbianUniCyr,serbianUniLat, and serbianLex. The examples for serbianLex present the full simulation for Experiment 1 of Baayen et al. (2011).

Key functionality for the user is provided by the functions orthoCoding, estimateWeights, and estimateActivations. orthoCoding calculates the letter n-grams for character strings, to be used as cues. It is assumed that meaning or meanings (separated by underscores if there are more then one) are available as outcomes. The frequency with which each (unique) combination of cues and outcomes occurs are required. For some example input data sets, see: danks, plurals, serbian, serbianUniCyr and serbianUniLat.

The function estimateWeights estimates the association strengths of cues to outcomes, using the equilibrium equations presented in Danks (2003). The function estimateActivations estimates the activations of outcomes (meanings) given cues (n-grams).

The Rcpp-based learn and learnLegacy functions use a C++ function to compute the conditional co-occurrence matrices required in the equilibrium equations. These are internally used by estimateWeights and should not be used directly by users of the package.

The key function for naive discriminative classification is ndlClassify; see data sets think and dative for examples.

References

Baayen, R. H. and Milin, P. and Filipovic Durdevic, D. and Hendrix, P. and Marelli, M., An amorphous model for morphological processing in visual comprehension based on naive discriminative learning. Psychological Review, 118, 438-482.

Baayen, R. H. (2011) Corpus linguistics and naive discriminative learning. Brazilian Journal of Applied Linguistics, 11, 295-328.

Arppe, A. and Baayen, R. H. (in prep.) Statistical classification and principles of human learning.

Examples

# NOT RUN {
# Rescorla-Wagner
data(lexample)

lexample$Cues <- orthoCoding(lexample$Word, grams=1)
lexample.rw <- RescorlaWagner(lexample, nruns=25, traceCue="h",
   traceOutcome="hand")
plot(lexample.rw)
mtext("h - hand", 3, 1)

data(numbers)

traceCues <- c( "exactly1", "exactly2", "exactly3", "exactly4", "exactly5",
   "exactly6", "exactly7", "exactly10", "exactly15")
traceOutcomes <- c("1", "2", "3", "4", "5", "6", "7", "10", "15")

ylimit <- c(0,1)
par(mfrow=c(3,3), mar=c(4,4,1,1))

for (i in 1:length(traceCues)) {
  numbers.rw <- RescorlaWagner(numbers, nruns=1, traceCue=traceCues[i],
     traceOutcome=traceOutcomes[i])
  plot(numbers.rw, ylimit=ylimit)
  mtext(paste(traceCues[i], " - ", traceOutcomes[i], sep=""), side=3, line=-1,
    cex=0.7)
}
par(mfrow=c(1,1))

# naive discriminative learning (for complete example, see serbianLex)
# This function uses a Unicode dataset.
data(serbianUniCyr)
serbianUniCyr$Cues <- orthoCoding(serbianUniCyr$WordForm, grams=2)
serbianUniCyr$Outcomes <- serbianUniCyr$LemmaCase
sw <- estimateWeights(cuesOutcomes=serbianUniCyr,hasUnicode=T)

desiredItems <- unique(serbianUniCyr["Cues"])
desiredItems$Outcomes=""
activations <- estimateActivations(desiredItems, sw)$activationMatrix
rownames(activations) <- unique(serbianUniCyr[["WordForm"]])

syntax <- c("acc", "dat", "gen", "ins", "loc", "nom", "Pl",  "Sg") 
activations2 <- activations[,!is.element(colnames(activations), syntax)]
head(rownames(activations2),50)
head(colnames(activations2),8)

image(activations2, xlab="word forms", ylab="meanings", xaxt="n", yaxt="n")
mtext(c("yena", "...", "zvuke"), side=1, line=1, at=c(0, 0.5, 1),  adj=c(0,0,1))
mtext(c("yena", "...", "zvuk"), side=2, line=1, at=c(0, 0.5, 1),   adj=c(0,0,1))

# naive discriminative classification
data(think)
think.ndl <- ndlClassify(Lexeme ~ Person + Number + Agent + Patient + Register,
   data=think)
summary(think.ndl)
plot(think.ndl, values="weights", type="hist", panes="multiple")
plot(think.ndl, values="probabilities", type="density")
# }