qlcMatrix (version 0.9.8)

splitText: Construct sparse matrices from parallel texts

Description

Convenience functions to read parallel texts and to split parallel texts into sparse matrices.

Usage

splitText(text, globalSentenceID = NULL, localSentenceID = names(text), sep = " ",
	simplify = FALSE, lowercase = TRUE)
read.text(file)

Value

When simplify = F, a list is returned with the following elements:

runningWords

single vector with complete text (ignoring original sentence breaks), separated into strings according to sep

wordforms

vector with all wordforms as attested in the text (according to the specified separator). Ordering of wordforms is done by ttMatrix, which by default uses the "C" collation locale.

lowercase

only returned when lowercase = T. Vector with all unique wordforms after conversion to lowercase.

RS

Sparse pattern matrix of class ngCMatrix with runningWords (R) as rows and sentence IDs (S) as columns. When globalSentenceID = NULL, then the sentences are the elements of the original text. Else, they are the specified globalSentenceIDs.

WR

Sparse pattern matrix of class ngCMatrix with wordforms (W) as rows and running words (R) as columns.

wW

only returned when lowercase = T. Sparse pattern matrix of class ngCMatrix linking between lowercased wordforms and original wordforms.

When simplify = T the result is a single sparse Matrix (of type dgCMatrix) linking wordforms (either with or without case) to sentences (either global or local). Note that the result with options (simplify = T, lowercase = F) will result in the sparse matrix as available at paralleltext.info (there the matrix is in .mtx format), with the wordforms included into the matrix as row names. However, note that the resulting matrix from the code here will include frequencies for words that occur more than once per sentence. These have been removed for the .mtx version available online.

Arguments

text

vector of strings, typically sentences with wordforms separated by space (see sep below). names of the vector elements are typically IDs to link across texts (cf. the format as used in bibles).

globalSentenceID

Vector of all IDs that might possibly occur in the parallel texts, used to parallelize the texts. Can for example be constructed by using union on the localSentenceIDs.

localSentenceID

Vector of the IDs for the actual sentences in the present text. Typically present as names of the text.

sep

Separator on which the sentences should be parsed into wordforms. The implementation is very simple here, there are no advanced options for guessing punctuation. The variation in punctuation across a wide variety of languages and scripts normally turns out to be too large to be easily automatically parsed. Any advanced parsing has to be done externally, and here simply the parsed symbol is used to actually split the text into parts. Typically, this parsing of sentences into wordforms will be performed using space sep = " ". See also bibles for some examples of such pre-parsing.

simplify

By default (when simplify = F), this function returns a list of objects that represent the encoding of the text into sparse matrices. With simplify = T this list is reduced to a single matrix (wordforms x globalSentenceID), with the actual wordforms as row names.

lowercase

By default, a mapping between the text and a lowercase version of the same text. In the default output (with simplify = F), this is a sparse matrix linking strings with mixed upper/lower case to string with only lower case. Note that case folding is locale-specific, but here a simple universal case-folding is used (as available through tolower).

file

file name (or full path) for a file to be read.

Author

Michael Cysouw

Details

The function splitText is actually just a nice examples of how pwMatrix, jMatrix, and ttMatrix can be used to work with parallel texts.

The function read.text is a convenience function to read parallel texts.

See Also

bibles for some texts that led to the development of this function. sim.words for a convenience function to easily extract possible translations equivalents through co-occurrence (using splitText for the data-preparation.)

Examples

# a trivial examples to see the results of this function:
text <- c("This is a sentence .","A sentence this is !","Is this a sentence ?")
splitText(text)
splitText(text, simplify = TRUE, lowercase = FALSE)

# reasonably quick with complete bibles (about 1-2 second per complete bible)
# texts with only New Testament is even quicker
data(bibles)
system.time(eng <- splitText(bibles$eng, bibles$verses))
system.time(deu <- splitText(bibles$deu, bibles$verses))

# Use example: Number of co-ocurrences between two bibles
# (this is more conveniently performed by the function sim.words)
# How often do words from the one language cooccur with words from the other language?
ENG <- (eng$wW * 1) %*% (eng$WR * 1) %*% (eng$RS * 1)
DEU <- (deu$wW * 1) %*% (deu$WR * 1) %*% (deu$RS * 1)
C <- tcrossprod(ENG,DEU)
rownames(C) <- eng$lowercase
colnames(C) <- deu$lowercase
C[	c("father","father's","son","son's"),
	c("vater","vaters","sohn","sohne","sohnes","sohns")
	]

# Pure counts are not very interesting. This is better:
R <- assocSparse(t(ENG), t(DEU))
rownames(R) <- eng$lowercase
colnames(R) <- deu$lowercase
R[	c("father","father's","son","son's"),
	c("vater","vaters","sohn","sohne","sohnes","sohns")
	]

# For example: best co-occurrences for the english word "mine"
sort(R["mine",], decreasing = TRUE)[1:10]

# \donttest{
# To get a quick-and-dirty translation matrix:
# adding maxima from both sides work quite well
# but this takes some time

cm <- colMax(R, which = TRUE, ignore.zero = FALSE)$which
rm <- rowMax(R, which = TRUE, ignore.zero = FALSE)$which
best <- cm + rm
best <- as(best, "nMatrix")

which(best["your",])
which(best["went",])

# A final speed check:
# split all 4 texts, and simplify them into one matrix
# They have all the same columns, so they can be rbind
system.time(all <- sapply(bibles[-1], function(x){splitText(x, bibles$verses, simplify = TRUE)}))
all <- do.call(rbind, all)

# then try a single co-occerrence measure on all pairs from these 72K words
# (so we are doing about 2.6e9 comparisons here!)
system.time( S <- cosSparse(t(all)) )

# this goes extremely fast! As long as everything fits into RAM this works nicely.
# Note that S quickly gets large
print(object.size(S), units = "auto")

# but most of it can be thrown away, because it is too low anyway
# this leads to a factor 10 reduction in size:
S <- drop0(S, tol = 0.2)
print(object.size(S), units = "auto")
# }