Files
b2txt25/language_model/srilm-1.7.3/man/html/training-scripts.1.html
2025-07-02 12:18:09 -07:00

406 lines
15 KiB
HTML

<! $Id: training-scripts.1,v 1.29 2019/09/09 22:35:37 stolcke Exp $>
<HTML>
<HEADER>
<TITLE>training-scripts</TITLE>
<BODY>
<H1>training-scripts</H1>
<H2> NAME </H2>
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
<H2> SYNOPSIS </H2>
<PRE>
<B>get-gt-counts</B> <B>max=</B><I>K</I> <B>out=</B><I>name</I> [ <I>counts</I> ... ] <B>&gt;</B> <I>gtcounts</I>
<B>make-abs-discount</B> <I>gtcounts</I>
<B>make-gt-discounts</B> <B>min=</B><I>min</I> <B>max=</B><I>max</I> <I>gtcounts</I>
<B>make-kn-counts</B> <B>order=</B><I>N</I> <B>max_per_file=</B><I>M</I> <B>output=</B><I>file</I> \
[ <B>no_max_order=1</B> ] <B>&gt;</B> <I>counts</I>
<B>make-kn-discounts</B> <B>min=</B><I>min</I> <I>gtcounts</I>
<B>make-batch-counts</B> <I>file-list</I> \
[ <I>batch-size</I> [ <I>filter</I> [ <I>count-dir</I> [ <I>options</I> ... ] ] ] ]
<B>merge-batch-counts</B> [ <B>-float-counts</B> ] [ <B>-l</B> <I>N</I> ] <I>count-dir</I> [ <I>file-list</I>|<I>start-iter</I> ]
<B>make-google-ngrams</B> [ <B>dir=</B><I>DIR</I> ] [ <B>per_file=</B><I>N</I> ] [ <B>gzip=0</B> ] \
[ <B>yahoo=1</B> ] [ <I>counts-file</I> ... ]
<B>continuous-ngram-count</B> [ <B>order=</B><I>N</I> ] [ <I>textfile</I> ... ]
<B>tolower-ngram-counts</B> [ <I>counts-file</I> ... ]
<B>uniq-ngram-counts</B> [ <I>counts-file</I> ... ]
<B>reverse-ngram-counts</B> [ <I>counts-file</I> ... ]
<B>reverse-text</B> [ <I>textfile</I> ... ]
<B>split-tagged-ngrams</B> [ <B>separator=</B><I>S</I> ] [ <I>counts-file</I> ... ]
<B>make-big-lm</B> <B>-name</B> <I>name</I> <B>-read</B> <I>counts</I> <B>-lm</B> <I>new-model</I> \
[ <B>-trust-totals</B> ] [ <B>-max-per-file</B> <I>M</I> ] [ <B>-ngram-filter</B> <I>filter</I> ] \
[ \fB-text <I>file</I> ] [ <I>ngram-options</I> ... ]
<B>replace-unk-words</B> <B>vocab=</B><I>vocab</I> [ <I>textfile</I> ... ]
<B>replace-words-with-classes</B> <B>classes=</B><I>classes</I> [ <B>outfile=</B><I>counts</I> ] \
[ <B>normalize=0</B>|<B>1</B> ] [ <B>addone=</B><I>K</I> ] [ <B>have_counts=1</B> ] [ <B>partial=1</B> ] \
[ <I>textfile</I> ... ]
<B>uniform-classes</B> <I>classes</I> <B>&gt;</B> <I>new-classes</I>
<B>make-diacritic-map</B> <I>vocab</I>
<B>vp2text</B> [ <I>textfile</I> ... ]
<B>compute-oov-rate</B> <I>vocab</I> [ <I>counts</I> ... ]
</PRE>
<H2> DESCRIPTION </H2>
These scripts perform convenience tasks associated with the training of
language models.
They complement and extend the basic N-gram model estimator in
<A HREF="ngram-count.1.html">ngram-count(1)</A>.
<P>
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
<A HREF="gunzip.1.html">gunzip(1)</A>
or
<A HREF="gzip.1.html">gzip(1)</A>
on the command line.
<P>
Also note that many of the scripts take their options with the
<A HREF="gawk.1.html">gawk(1)</A>
syntax
<I>option</I><B>=</B><I>value</I><B></B><I></I><B></B><I></I>
instead of the more common
<B>-</B><I>option</I><B></B><I></I><B></B><I></I><B></B>
<I>value</I>.<I></I><I></I><I></I>
<P>
<B> get-gt-counts </B>
computes the counts-of-counts statistics needed in Good-Turing smoothing.
The frequencies of counts up to
<I> K </I>
are computed (default is 10).
The results are stored in a series of files with root
<I>name</I>,<I></I><I></I><I></I>
<B><I>name</I>.gt1counts</B>,<B></B><B></B><B></B>
<B><I>name</I>.gt2counts</B>,<B></B><B></B><B></B>
...,
<B><I>name</I>.gt<I>N</I>counts</B>.<B></B><B></B><B></B>
It is assumed that the input counts have been properly merged, i.e.,
that there are no duplicated N-grams.
<P>
<B> make-gt-discounts </B>
takes one of the output files of
<B> get-gt-counts </B>
and computes the corresponding Good-Turing discounting factors.
The output can then be passed to
<A HREF="ngram-count.1.html">ngram-count(1)</A>
via the
<B>-gt</B><I>n</I><B></B><I></I><B></B><I></I><B></B>
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,
<B>get-gt-counts</B>/<B>make-gt-discounts</B><B></B><B></B>
can process arbitrarily large count files, since they do not need to
read the counts into memory (unlike
<B>ngram-count</B>).<B></B><B></B><B></B>
<P>
<B> make-abs-discount </B>
computes the absolute discounting constant needed for the
<B> ngram-count </B>
<B>-cdiscount</B><I>n</I><B></B><I></I><B></B><I></I><B></B>
options.
Input is one of the files produced by
<B>get-gt-counts</B>.<B></B><B></B><B></B>
<P>
<B> make-kn-discount </B>
computes the discounting constants used by the modified Kneser-Ney
smoothing method.
Input is one of the files produced by
<B>get-gt-counts</B>.<B></B><B></B><B></B>
This script also implements a method for extrapolating missing
counts of counts as described in
Wang et al. (2007).
<P>
<B> make-batch-counts </B>
performs the first stage in the construction of very large N-gram count
files.
<I> file-list </I>
is a list of input text files.
Lines starting with a `#' character are ignored.
These files will be grouped into batches of size
<I> batch-size </I>
(default 10)
that are then processed in one run of
<B> ngram-count </B>
each.
For maximum performance,
<I> batch-size </I>
should be as large as possible without triggering paging.
Optionally, a
<I> filter </I>
script or program can be given to condition the input texts.
The N-gram count files are left in directory
<I> count-dir </I>
(``counts'' by default), where they can be found by a subsequent
run of
<B>merge-batch-counts</B>.<B></B><B></B><B></B>
All following
<I> options </I>
are passed to
<B>ngram-count</B>,<B></B><B></B><B></B>
e.g., to control N-gram order, vocabulary, etc.
(no options triggering model estimation should be included).
<P>
<B> merge-batch-counts </B>
completes the construction of large count files by merging the
batched counts left in
<I> count-dir </I>
until a single count file is produced.
Optionally, a
<I> file-list </I>
of count files to combine can be specified; otherwise all count files
in
<I> count-dir </I>
from a prior run of
<B> make-batch-counts </B>
will be merged.
A number as second argument restarts the merging process at iteration
<I>start-iter</I>.<I></I><I></I><I></I>
This is convenient if merging fails to complete for some reason
(e.g., for temporary lack of disk space).
The
<B> -float-counts </B>
option should be specific if the counts are real-valued.
The
<B> -l </B>
option specifies the number of files to merge in each iteration;
the default is 2.
<P>
<B> make-google-ngrams </B>
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
<A HREF="ngram-count.1.html">ngram-count(1)</A>
<B> -read-google </B>
option.
Optional arguments specify the output directory
<I> dir </I>
and the size
<I> N </I>
of individual N-gram files
(default is 10 million N-grams per file).
The
<B> gzip=0 </B>
option writes plain, as opposed to compressed, files.
The
<B> yahoo=1 </B>
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.
<P>
<B> continuous-ngram-count </B>
generates N-grams that span line breaks (which are usually taken to
be sentence boundaries).
To count N-grams across line breaks use
<PRE>
continuous-ngram-count <I>textfile</I> | ngram-count -read -
</PRE>
The argument
<I> N </I>
controls the order of N-grams counted (default 3), and
should match the argument of
<B> ngram-count </B>
<B>-order</B>.<B></B><B></B><B></B>
<P>
<B> tolower-ngram-counts </B>
maps an N-gram counts file to all-lowercase.
No merging of N-grams that become identical in the process is done.
<P>
<B> uniq-ngram-counts </B>
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.,
<BR>
tolower-ngram-counts <I>INPUT</I> | sort | uniq-ngram-counts &gt; \fiOUTPUT\fP
<BR>
would do much of the same thing as
<BR>
ngram-counts -read <I>INPUT</I> -tolower -sort -write <I>OUTPUT</I>
<BR>
but in a more memory-efficient manner (without reading all counts into memory).
<P>
<B> reverse-ngram-counts </B>
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
<A HREF="ngram-count.1.html">ngram-count(1)</A>),
use
<BR>
reverse-ngram-counts <I>count-file</I> | \
<BR>
ngram-count -read - -recompute -write - | \
<BR>
reverse-ngram-counts &gt; <I>new-counts</I>
<BR>
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.
<P>
<B> reverse-text </B>
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
<B> ngram </B>
<B>-reverse</B><B></B><B></B><B></B>
option.
<P>
<B> split-tagged-ngrams </B>
expands N-gram count of word/tag pairs into mixed N-grams
of words and tags.
The optional
<B>separator=</B><I>S</I><B></B><I></I><B></B><I></I><B></B>
argument allows the delimiting character, which defaults to "/",
to be modified.
<P>
<B> make-big-lm </B>
constructs large N-gram models in a more memory-efficient way than
<B> ngram-count </B>
by itself.
It does so by precomputing the Good-Turing or Kneser-Ney smoothing parameters
from the full set of counts, and then instructing
<B> ngram-count </B>
to store only a subset of the counts in memory,
namely those of N-grams to be retained in the model.
The
<I> name </I>
parameter is used to name various auxiliary files.
<I> counts </I>
contains the raw N-gram counts; it may be (and usually is) a compressed file.
Unlike with
<B>ngram-count</B>,<B></B><B></B><B></B>
the
<B> -read </B>
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
<B> -trust-totals </B>
options may be given for efficiency.
The
<B> -text </B>
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
<I> options </I>
are passed to
<B> ngram-count </B>
(only options affecting model estimation should be given).
Smoothing methods other than Good-Turing, modified Kneser-Ney and Witten-Bell are not
supported by
<B>make-big-lm</B>.<B></B><B></B><B></B>
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
<B> merge-batch-counts </B>
command, and the
<B> -max-per-file </B>
option controls how many counts are to be stored per batch, and
should be chosen so that these batches fit in real memory.
The
<B> -ngram-filter </B>
option allows specification of a command through which the input N-gram
counts are piped, e.g., to convert from some non-standard format.
<P>
<B> make-kn-counts </B>
computes the modified lower-order counts used by the KN smoothing method.
It is invoked as a helper scripts by
<B> make-big-lm . </B>
<P>
<B> replace-unk-words </B>
replaces words not appearing in the
<I> vocab </I>
file with the unknown word tag
<B>&lt;unk&gt;</B>.<B></B><B></B><B></B>
This is useful for preparing text data for LM training.
Only the first token on each line in the
<I> vocab </I>
file is significant, so both word lists and unigram count files may be used.
<P>
<B> replace-words-with-classes </B>
replaces expansions of word classes with the corresponding class labels.
<I> classes </I>
specifies class expansions in
<A HREF="classes-format.5.html">classes-format(5)</A>.
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
<I> counts </I>
will receive the expansion counts resulting from the replacements.
<B> normalize=0 </B>
or
<B> 1 </B>
indicates whether the counts should be normalized to probabilities
(default is 1).
The
<B> addone </B>
option may be used to smooth the expansion probabilities by adding
<I> K </I>
to each count (default 1).
The option
<B> have_counts=1 </B>
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
<A HREF="ngram-count.1.html">ngram-count(1)</A>
reads count data.
The option
<B> partial=1 </B>
prevents multi-word class expansions from being replaced when more than
one space character occurs inbetween the words.
<P>
<B> uniform-classes </B>
takes a file in
<A HREF="classes-format.5.html">classes-format(5)</A>
and adds uniform probabilities to expansions that don't have a probability
explicitly stated.
<P>
<B> make-diacritic-map </B>
constructs a map file that pairs an ASCII-fied version of the words in
<I> vocab </I>
with all the occurring non-ASCII word forms.
Such a map file can then be used with
<A HREF="disambig.1.html">disambig(1)</A>
and a language model
to reconstruct the non-ASCII word form with diacritics from an ASCII
text.
<P>
<B> vp2text </B>
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.
<P>
<B> compute-oov-rate </B>
determines the out-of-vocabulary rate of a corpus from its unigram
<I> counts </I>
and a target vocabulary list in
<I>vocab</I>.<I></I><I></I><I></I>
<H2> SEE ALSO </H2>
<A HREF="ngram-count.1.html">ngram-count(1)</A>, <A HREF="ngram.1.html">ngram(1)</A>, <A HREF="classes-format.5.html">classes-format(5)</A>, <A HREF="disambig.1.html">disambig(1)</A>, <A HREF="select-vocab.1.html">select-vocab(1)</A>.
<BR>
W. Wang, A. Stolcke, J. Zheng, ``Reranking machine translation hypotheses with structured
and web-based language models'', <I>Proc. IEEE ASRU Workshop</I>, pp. 159-164, 2007.
<H2> BUGS </H2>
Some of the tools could be generalized and/or made more robust to
misuse.
<BR>
Several of these tools are gawk scripts and depending on prevailing locale
settings might require an LC_NUMERIC=C environment variable.
<H2> AUTHOR </H2>
Andreas Stolcke &lt;stolcke@icsi.berkeley.edu&gt;
<BR>
Copyright (c) 1995-2008 SRI International
<BR>
Copyright (c) 2013-2017 Andreas Stolcke
<BR>
Copyright (c) 2013-2017 Microsoft Corp.
</BODY>
</HTML>