training-scripts

training-scripts

NAME

training-scripts, compute-oov-rate, continuous-ngram-count, get-gt-counts, make-abs-discount, make-batch-counts, make-big-lm, make-diacritic-map, make-google-ngrams, make-gt-discounts, make-kn-counts, make-kn-discounts, merge-batch-counts, replace-unk-words, replace-words-with-classes, reverse-ngram-counts, split-tagged-ngrams, reverse-text, tolower-ngram-counts, uniform-classes, uniq-ngram-counts, vp2text - miscellaneous conveniences for language model training

SYNOPSIS

get-gt-counts max=K out=name [ counts ... ] > gtcounts
make-abs-discount gtcounts
make-gt-discounts min=min max=max gtcounts
make-kn-counts order=N max_per_file=M output=file \
	[ no_max_order=1 ] > counts
make-kn-discounts min=min gtcounts
make-batch-counts file-list \
	[ batch-size [ filter [ count-dir [ options ... ] ] ] ]
merge-batch-counts [ -float-counts ] [ -l N ] count-dir [ file-list|start-iter ]
make-google-ngrams [ dir=DIR ] [ per_file=N ] [ gzip=0 ] \
	[ yahoo=1 ] [ counts-file ... ]
continuous-ngram-count [ order=N ] [ textfile ... ]
tolower-ngram-counts [ counts-file ... ]
uniq-ngram-counts [ counts-file ... ]
reverse-ngram-counts [ counts-file ... ]
reverse-text [ textfile ... ]
split-tagged-ngrams [ separator=S ] [ counts-file ... ]
make-big-lm -name name -read counts -lm new-model \
	[ -trust-totals ] [ -max-per-file M ] [ -ngram-filter filter ] \
	[ \fB-text file ] [ ngram-options ... ]
replace-unk-words vocab=vocab [ textfile ... ]
replace-words-with-classes classes=classes [ outfile=counts ] \
	[ normalize=0|1 ] [ addone=K ] [ have_counts=1 ] [ partial=1 ] \
	[ textfile ... ]
uniform-classes classes > new-classes
make-diacritic-map vocab
vp2text [ textfile ... ]
compute-oov-rate vocab [ counts ... ]

DESCRIPTION

These scripts perform convenience tasks associated with the training of language models. They complement and extend the basic N-gram model estimator in ngram-count(1).

Since these tools are implemented as scripts they don't automatically input or output compressed data files correctly, unlike the main SRILM tools. However, since most scripts work with data from standard input or to standard output (by leaving out the file argument, or specifying it as ``-'') it is easy to combine them with gunzip(1) or gzip(1) on the command line.

Also note that many of the scripts take their options with the gawk(1) syntax option=value instead of the more common -option value.

get-gt-counts computes the counts-of-counts statistics needed in Good-Turing smoothing. The frequencies of counts up to K are computed (default is 10). The results are stored in a series of files with root name, name.gt1counts, name.gt2counts, ..., name.gtNcounts. It is assumed that the input counts have been properly merged, i.e., that there are no duplicated N-grams.

make-gt-discounts takes one of the output files of get-gt-counts and computes the corresponding Good-Turing discounting factors. The output can then be passed to ngram-count(1) via the -gtn options to control the smoothing during model estimation. Precomputing the GT discounting in this fashion has the advantage that the GT statistics are not affected by restricting N-grams to a limited vocabulary. Also, get-gt-counts/make-gt-discounts can process arbitrarily large count files, since they do not need to read the counts into memory (unlike ngram-count).

make-abs-discount computes the absolute discounting constant needed for the ngram-count -cdiscountn options. Input is one of the files produced by get-gt-counts.

make-kn-discount computes the discounting constants used by the modified Kneser-Ney smoothing method. Input is one of the files produced by get-gt-counts. This script also implements a method for extrapolating missing counts of counts as described in Wang et al. (2007).

make-batch-counts performs the first stage in the construction of very large N-gram count files. file-list is a list of input text files. Lines starting with a `#' character are ignored. These files will be grouped into batches of size batch-size (default 10) that are then processed in one run of ngram-count each. For maximum performance, batch-size should be as large as possible without triggering paging. Optionally, a filter script or program can be given to condition the input texts. The N-gram count files are left in directory count-dir (``counts'' by default), where they can be found by a subsequent run of merge-batch-counts. All following options are passed to ngram-count, e.g., to control N-gram order, vocabulary, etc. (no options triggering model estimation should be included).

merge-batch-counts completes the construction of large count files by merging the batched counts left in count-dir until a single count file is produced. Optionally, a file-list of count files to combine can be specified; otherwise all count files in count-dir from a prior run of make-batch-counts will be merged. A number as second argument restarts the merging process at iteration start-iter. This is convenient if merging fails to complete for some reason (e.g., for temporary lack of disk space). The -float-counts option should be specific if the counts are real-valued. The -l option specifies the number of files to merge in each iteration; the default is 2.

make-google-ngrams takes a sorted count file as input and creates an indexed directory structure, in a format developed by Google to store very large N-gram collections. The resulting directory can then be used with the ngram-count(1) -read-google option. Optional arguments specify the output directory dir and the size N of individual N-gram files (default is 10 million N-grams per file). The gzip=0 option writes plain, as opposed to compressed, files. The yahoo=1 option may be used to read N-gram count files in Yahoo-GALE format. Note that the count files have to first be sorted lexicographically in a separate invocation of sort.

continuous-ngram-count generates N-grams that span line breaks (which are usually taken to be sentence boundaries). To count N-grams across line breaks use

	continuous-ngram-count textfile | ngram-count -read -
The argument N controls the order of N-grams counted (default 3), and should match the argument of ngram-count -order.

tolower-ngram-counts maps an N-gram counts file to all-lowercase. No merging of N-grams that become identical in the process is done.

uniq-ngram-counts combines successive counts that pertain to the same N-gram into a single line. This only affects repeated N-grams that appear on successive lines, so a prior sort command is typically used, e.g.,
tolower-ngram-counts INPUT | sort | uniq-ngram-counts > \fiOUTPUT\fP
would do much of the same thing as
ngram-counts -read INPUT -tolower -sort -write OUTPUT
but in a more memory-efficient manner (without reading all counts into memory).

reverse-ngram-counts reverses the word order of N-grams in a counts file or stream. For example, to recompute lower-order counts from higher-order ones, but do the summation over preceding words (rather than following words, as in ngram-count(1)), use
reverse-ngram-counts count-file | \
ngram-count -read - -recompute -write - | \
reverse-ngram-counts > new-counts
Also, start-sentence tags are replaced with end-sentence tags, and vice-versa, so that reverse-direction LMs can be trained from forward-direction N-gram counts.

reverse-text reverses the word order in text files, line-by-line. Start- and end-sentence tags, if present, will be preserved. This reversal is appropriate for preprocessing training data for LMs that are meant to be used with the ngram -reverse option.

split-tagged-ngrams expands N-gram count of word/tag pairs into mixed N-grams of words and tags. The optional separator=S argument allows the delimiting character, which defaults to "/", to be modified.

make-big-lm constructs large N-gram models in a more memory-efficient way than ngram-count by itself. It does so by precomputing the Good-Turing or Kneser-Ney smoothing parameters from the full set of counts, and then instructing ngram-count to store only a subset of the counts in memory, namely those of N-grams to be retained in the model. The name parameter is used to name various auxiliary files. counts contains the raw N-gram counts; it may be (and usually is) a compressed file. Unlike with ngram-count, the -read option can be repeated to concatenate multiple count files, but the arguments must be regular files; reading from stdin is not supported. If Good-Turing smoothing is used and the file contains complete lower-order counts corresponding to the sums of higher-order counts, then the -trust-totals options may be given for efficiency. The -text option specifies a test set to which the LM is to be applied, and builds the LM in such a way that only N-gram context occurring in the test data are included in the model, this saving space at the expense of generality. All other options are passed to ngram-count (only options affecting model estimation should be given). Smoothing methods other than Good-Turing, modified Kneser-Ney and Witten-Bell are not supported by make-big-lm. Kneser-Ney smoothing also requires enough disk space to compute and store the modified lower-order counts used by the KN method. This is done using the merge-batch-counts command, and the -max-per-file option controls how many counts are to be stored per batch, and should be chosen so that these batches fit in real memory. The -ngram-filter option allows specification of a command through which the input N-gram counts are piped, e.g., to convert from some non-standard format.

make-kn-counts computes the modified lower-order counts used by the KN smoothing method. It is invoked as a helper scripts by make-big-lm .

replace-unk-words replaces words not appearing in the vocab file with the unknown word tag <unk>. This is useful for preparing text data for LM training. Only the first token on each line in the vocab file is significant, so both word lists and unigram count files may be used.

replace-words-with-classes replaces expansions of word classes with the corresponding class labels. classes specifies class expansions in classes-format(5). Substitutions are performed at each word position in left to right order, with the longest matching right-hand-side of any class expansion. If several classes match a pseudo-random choice is made. Optionally, the file counts will receive the expansion counts resulting from the replacements. normalize=0 or 1 indicates whether the counts should be normalized to probabilities (default is 1). The addone option may be used to smooth the expansion probabilities by adding K to each count (default 1). The option have_counts=1 indicates that the input consists of N-gram counts and that replacement should be performed on them. Note this will not merge counts that have been mapped to identical N-grams, since this is done automatically when ngram-count(1) reads count data. The option partial=1 prevents multi-word class expansions from being replaced when more than one space character occurs inbetween the words.

uniform-classes takes a file in classes-format(5) and adds uniform probabilities to expansions that don't have a probability explicitly stated.

make-diacritic-map constructs a map file that pairs an ASCII-fied version of the words in vocab with all the occurring non-ASCII word forms. Such a map file can then be used with disambig(1) and a language model to reconstruct the non-ASCII word form with diacritics from an ASCII text.

vp2text is a reimplementation of the filter used in the DARPA Hub-3 and Hub-4 CSR evaluations to convert ``verbalized punctuation'' texts to language model training data.

compute-oov-rate determines the out-of-vocabulary rate of a corpus from its unigram counts and a target vocabulary list in vocab.

SEE ALSO

ngram-count(1), ngram(1), classes-format(5), disambig(1), select-vocab(1).
W. Wang, A. Stolcke, J. Zheng, ``Reranking machine translation hypotheses with structured and web-based language models'', Proc. IEEE ASRU Workshop, pp. 159-164, 2007.

BUGS

Some of the tools could be generalized and/or made more robust to misuse.
Several of these tools are gawk scripts and depending on prevailing locale settings might require an LC_NUMERIC=C environment variable.

AUTHOR

Andreas Stolcke <stolcke@icsi.berkeley.edu>
Copyright (c) 1995-2008 SRI International
Copyright (c) 2013-2017 Andreas Stolcke
Copyright (c) 2013-2017 Microsoft Corp.