616 lines
21 KiB
HTML
616 lines
21 KiB
HTML
|
<! $Id: nbest-optimize.1,v 1.31 2019/09/09 22:35:37 stolcke Exp $>
|
||
|
|
<HTML>
|
||
|
|
<HEADER>
|
||
|
|
<TITLE>nbest-optimize</TITLE>
|
||
|
|
<BODY>
|
||
|
|
<H1>nbest-optimize</H1>
|
||
|
|
<H2> NAME </H2>
|
||
|
|
nbest-optimize - optimize score combination for N-best word error minimization
|
||
|
|
<H2> SYNOPSIS </H2>
|
||
|
|
<PRE>
|
||
|
|
<B>nbest-optimize</B> [ <B>-help</B> ] <I>option</I> ... [ <I>scoredir</I> ... ]
|
||
|
|
</PRE>
|
||
|
|
<H2> DESCRIPTION </H2>
|
||
|
|
<B> nbest-optimize </B>
|
||
|
|
reads a set of N-best lists, additional score files, and corresponding
|
||
|
|
reference transcripts and optimizes the score combination weights
|
||
|
|
so as to minimize the word error of a classifier that performs
|
||
|
|
word-level posterior probability maximization.
|
||
|
|
The optimized weights are meant to be used with
|
||
|
|
<A HREF="nbest-lattice.1.html">nbest-lattice(1)</A>
|
||
|
|
and the
|
||
|
|
<B> -use-mesh </B>
|
||
|
|
option,
|
||
|
|
or the
|
||
|
|
<B> nbest-rover </B>
|
||
|
|
script (see
|
||
|
|
<A HREF="nbest-scripts.1.html">nbest-scripts(1)</A>).
|
||
|
|
<B> nbest-optimize </B>
|
||
|
|
determines both the best relative weighting of knowledge source scores
|
||
|
|
and the optimal
|
||
|
|
<B> -posterior-scale </B>
|
||
|
|
parameter that controls the peakedness of the posterior distribution.
|
||
|
|
<P>
|
||
|
|
Alternatively,
|
||
|
|
<B> nbest-optimize </B>
|
||
|
|
can also optimize weights for a standard, 1-best hypothesis rescoring that
|
||
|
|
selects entire (sentence) hypotheses
|
||
|
|
(<B>-1best</B><B></B><B></B>
|
||
|
|
option).
|
||
|
|
In this mode sentence-level error counts may be read from external files,
|
||
|
|
or computed on the fly from the reference strings.
|
||
|
|
The weights obtained are meant to be used for N-best list rescoring with
|
||
|
|
<B> rescore-reweight </B>
|
||
|
|
(see
|
||
|
|
<A HREF="nbest-scripts.1.html">nbest-scripts(1)</A>).
|
||
|
|
<P>
|
||
|
|
A third optimization criterion is the BLEU score used in machine translation.
|
||
|
|
This also requires the associated scores to be read from external files.
|
||
|
|
<P>
|
||
|
|
One of three optimization algorithms are available:
|
||
|
|
<DL>
|
||
|
|
<DT>1.
|
||
|
|
<DD>
|
||
|
|
The default optimization method is gradient descent on a smoothed (sigmoidal)
|
||
|
|
approximation of the true 0/1 word error function (Katagiri et al. 1990).
|
||
|
|
Therefore, the result can only be expected to be a
|
||
|
|
<I> local </I>
|
||
|
|
minimum of the error surface.
|
||
|
|
(A more global search can be attempted by specifying different starting
|
||
|
|
points.)
|
||
|
|
Another approximation is that the error function is computed assuming a fixed
|
||
|
|
multiple alignment of all N-best hypotheses and the reference string,
|
||
|
|
which tends to slightly overestimate the true pairwise error between any
|
||
|
|
single hypothesis and the reference.
|
||
|
|
<DT>2.
|
||
|
|
<DD>
|
||
|
|
An alternative search strategy uses a simplex-based "Amoeba" search on
|
||
|
|
the (non-smoothed) error function (Press et al. 1988).
|
||
|
|
The search is restarted multiple times to avoid local minima.
|
||
|
|
<DT>3.
|
||
|
|
<DD>
|
||
|
|
A third algorithm uses Powell search (Press et al. 1988)
|
||
|
|
on the (non-smoothed) error function.
|
||
|
|
</DD>
|
||
|
|
</DL>
|
||
|
|
<H2> OPTIONS </H2>
|
||
|
|
<P>
|
||
|
|
Each filename argument can be an ASCII file, or a
|
||
|
|
compressed file (name ending in .Z or .gz), or ``-'' to indicate
|
||
|
|
stdin/stdout.
|
||
|
|
<DL>
|
||
|
|
<DT><B> -help </B>
|
||
|
|
<DD>
|
||
|
|
Print option summary.
|
||
|
|
<DT><B> -version </B>
|
||
|
|
<DD>
|
||
|
|
Print version information.
|
||
|
|
<DT><B>-debug</B><I> level</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Controls the amount of output (the higher the
|
||
|
|
<I>level</I>,<I></I><I></I><I></I>
|
||
|
|
the more).
|
||
|
|
At level 1, error statistics at each iteration are printed.
|
||
|
|
At level 2, word alignments are printed.
|
||
|
|
At level 3, full score matrix is printed.
|
||
|
|
At level 4, detailed information about word hypothesis ranking is printed
|
||
|
|
for each training iteration and sample.
|
||
|
|
<DT><B>-nbest-files</B><I> file-list</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Specifies the set of N-best files as a list of filenames.
|
||
|
|
Three sets of standard scores are extracted from the N-best files:
|
||
|
|
the acoustic model score, the language model score, and the number of
|
||
|
|
words (for insertion penalty computation).
|
||
|
|
See
|
||
|
|
<A HREF="nbest-format.5.html">nbest-format(5)</A>
|
||
|
|
for details.
|
||
|
|
<BR>
|
||
|
|
In BLEU optimization mode, since there is no acoustic score, the
|
||
|
|
position
|
||
|
|
of the first score is taken by the "ac-replacement" score, which can be
|
||
|
|
any score used by the machine translation system.
|
||
|
|
A typical example is a score measuring
|
||
|
|
word order distortion between the source and target languages.
|
||
|
|
<DT><B>-xval-files</B><I> file-list</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Use the N-best files listed in
|
||
|
|
<I> file-list </I>
|
||
|
|
for cross-validation.
|
||
|
|
Optimization stops once no improvement is found on the cross-validation set
|
||
|
|
(while the direction of the search is obtained from the training set).
|
||
|
|
This is only supported with the gradient-descent and Amoeba optimization
|
||
|
|
methods.
|
||
|
|
<DT><B>-srinterp-format</B><I></I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Parse n-best list in SRInterp format, which has
|
||
|
|
features and text in the same line. n-best optimize will also generate
|
||
|
|
rover-control file in SRInterp format, where each line is in the form of:
|
||
|
|
<PRE>
|
||
|
|
<I>F1=V1</I> <I>F2=V2</I> ... <I>Fm=Vm</I> <I>W1</I> <I>W2</I> ...<I>Wn</I>
|
||
|
|
</PRE>
|
||
|
|
where
|
||
|
|
<I> F1 </I>
|
||
|
|
through
|
||
|
|
<I> Fm </I>
|
||
|
|
are feature names,
|
||
|
|
<I> V1 </I>
|
||
|
|
through
|
||
|
|
<I> Vm </I>
|
||
|
|
are feature values,
|
||
|
|
<I> W1 </I>
|
||
|
|
through
|
||
|
|
<I> Wn </I>
|
||
|
|
are words.
|
||
|
|
Also generate SRInterp control file, in the format of:
|
||
|
|
<PRE>
|
||
|
|
<I>F1:S1</I> <I>F2:S2</I> ... <I>Fm:Sm</I>
|
||
|
|
</PRE>
|
||
|
|
where
|
||
|
|
<I> S1 </I>
|
||
|
|
through
|
||
|
|
<I> Sm </I>
|
||
|
|
are scaling factors (weights) for feature
|
||
|
|
<I> F1 </I>
|
||
|
|
through
|
||
|
|
<I> Fm </I>
|
||
|
|
.
|
||
|
|
<DT><B>-refs</B><I> references</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Specifies the reference transcripts.
|
||
|
|
Each line in
|
||
|
|
<I> references </I>
|
||
|
|
must contain the sentence ID (the last component in the N-best filename
|
||
|
|
path, minus any suffixes) followed by zero or more reference words.
|
||
|
|
<DT><B>-insertion-weight</B><I> W</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Weight insertion errors by a factor
|
||
|
|
<I>W</I>.<I></I><I></I><I></I>
|
||
|
|
This may be useful to optimize for keyword spotting tasks where
|
||
|
|
insertions have a cost different from deletion and substitution errors.
|
||
|
|
<DT><B>-word-weights</B><I> file</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Read a table of words and weights from
|
||
|
|
<I>file</I>.<I></I><I></I><I></I>
|
||
|
|
Each word error is weighted according to the word-specific weight.
|
||
|
|
The default weight is 1, and used if a word has no specified weight.
|
||
|
|
Also, when this option is used, substitution errors are counted
|
||
|
|
as the sum of a deletion and an insertion error, as opposed to counting
|
||
|
|
as 1 error as in traditional word error computation.
|
||
|
|
<DT><B>-anti-refs</B><I> file</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Read a file of "anti-references" for use with the
|
||
|
|
<B> -anti-ref-weight </B>
|
||
|
|
option (see below).
|
||
|
|
<DT><B>-anti-ref-weight</B><I> W</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Compute the hypothesis errors for 1-best optimization by adding the
|
||
|
|
edit distance with respect to the "anti-references" times the weight
|
||
|
|
<I> W </I>
|
||
|
|
to the regular error count.
|
||
|
|
If
|
||
|
|
<I> W </I>
|
||
|
|
is negative this will tend to generate hypotheses that are different from
|
||
|
|
the anti-references (hence the name).
|
||
|
|
<DT><B> -1best </B>
|
||
|
|
<DD>
|
||
|
|
Select optimization for standard sentence-level hypothesis selection.
|
||
|
|
<DT><B> -1best-first </B>
|
||
|
|
<DD>
|
||
|
|
Optimized first using
|
||
|
|
<B> -1best </B>
|
||
|
|
mode, then switch to full optimization.
|
||
|
|
This is an effective way to quickly bring the score weights near an
|
||
|
|
optimal point, and then fine-tune them jointly with the posterior scale
|
||
|
|
parameter.
|
||
|
|
<DT><B>-errors</B><I> dir</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
In 1-best mode, optimize for error counts that are stored in separate files
|
||
|
|
in directory
|
||
|
|
<I>dir</I>.<I></I><I></I><I></I>
|
||
|
|
Each N-best list must have a matching error counts file of the same
|
||
|
|
basename in
|
||
|
|
<I>dir</I>.<I></I><I></I><I></I>
|
||
|
|
Each file contains 7 columns of numbers in the format
|
||
|
|
<PRE>
|
||
|
|
<I>wcr</I> <I>wer</I> <I>nsub</I> <I>ndel</I> <I>nins</I> <I>nerr</I> <I>nw</I>
|
||
|
|
</PRE>
|
||
|
|
Only the last two columns (number of errors and words, respectively) are used.
|
||
|
|
<BR>
|
||
|
|
If this option is omitted, errors will be computed from the N-best hypotheses
|
||
|
|
and the reference transcripts.
|
||
|
|
<DT><B>-bleu-counts</B><I> dir</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Perform BLEU optimization, reading BLEU reference counts from directory
|
||
|
|
<I>dir</I>.<I></I><I></I><I></I>
|
||
|
|
Each N-best list must have a matching counts file of the same
|
||
|
|
basename in
|
||
|
|
<I>dir</I>,<I></I><I></I><I></I>
|
||
|
|
containing the following information:
|
||
|
|
<PRE>
|
||
|
|
<I>N</I> <I>M</I> <I>L1</I> ... <I>LM</I>
|
||
|
|
</PRE>
|
||
|
|
where
|
||
|
|
<I> N </I>
|
||
|
|
is the number of hypotheses in the N-best list,
|
||
|
|
<I> M </I>
|
||
|
|
is the number of references for the utterance,
|
||
|
|
and
|
||
|
|
<I> L1 </I>
|
||
|
|
through
|
||
|
|
<I> LM </I>
|
||
|
|
are the reference lengths (word counts) for each reference.
|
||
|
|
Following this line, there are
|
||
|
|
<I> N </I>
|
||
|
|
lines of the form
|
||
|
|
<PRE>
|
||
|
|
<I>K</I> <I>C1</I> <I>C2</I> ... <I>Cm</I>
|
||
|
|
</PRE>
|
||
|
|
where
|
||
|
|
<I> K </I>
|
||
|
|
is the number of words in the hypothesis and
|
||
|
|
<I> C1 </I>
|
||
|
|
through
|
||
|
|
<I> Cm </I>
|
||
|
|
are the N-gram counts occurring in the references for each N-gram order
|
||
|
|
1, ...,
|
||
|
|
<I>m</I>.<I></I><I></I><I></I>
|
||
|
|
Currently,
|
||
|
|
<I> m </I>
|
||
|
|
is limited to 4.
|
||
|
|
<DT><B> -minimum-bleu-reference </B>
|
||
|
|
<DD>
|
||
|
|
Use shortest reference length to compute the BLEU brevity penalty.
|
||
|
|
<DT><B> -closest-bleu-reference </B>
|
||
|
|
<DD>
|
||
|
|
Use closest reference length for each translation hypothesis to compute
|
||
|
|
the BLEU brevity penalty.
|
||
|
|
<DT><B> -average-bleu-reference </B>
|
||
|
|
<DD>
|
||
|
|
Use average reference length to compute the BLEU brevity penalty.
|
||
|
|
<DT><B> -error-bleu-ratio R </B>
|
||
|
|
<DD>
|
||
|
|
Specifies the weight of error rate when combined with BLEU as optimization
|
||
|
|
objective: <I>(1-BLEU) + ERR x R</I>.
|
||
|
|
<I> ERR </I>
|
||
|
|
is error rate computed by #errors/#references.
|
||
|
|
<DT><B>-max-nbest</B><I> n</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Limits the number of hypotheses read from each N-best list to the first
|
||
|
|
<I>n</I>.<I></I><I></I><I></I>
|
||
|
|
<DT><B>-rescore-lmw</B><I> lmw</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Sets the language model weight used in combining the language model log
|
||
|
|
probabilities with acoustic log probabilities.
|
||
|
|
This is used to compute initial aggregate hypotheses scores.
|
||
|
|
<DT><B>-rescore-wtw</B><I> wtw</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Sets the word transition weight used to weight the number of words relative to
|
||
|
|
the acoustic log probabilities.
|
||
|
|
This is used to compute initial aggregate hypotheses scores.
|
||
|
|
<DT><B>-posterior-scale</B><I> scale</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Initial value for scaling log posteriors.
|
||
|
|
The total weighted log score is divided by
|
||
|
|
<I> scale </I>
|
||
|
|
when computing normalized posterior probabilities.
|
||
|
|
This controls the peakedness of the posterior distribution.
|
||
|
|
The default value is whatever was chosen for
|
||
|
|
<B>-rescore-lmw</B>,<B></B><B></B><B></B>
|
||
|
|
so that language model scores are scaled to have weight 1,
|
||
|
|
and acoustic scores have weight 1/<I>lmw</I>.
|
||
|
|
<DT><B> -combine-linear </B>
|
||
|
|
<DD>
|
||
|
|
Compute aggregate scores by linear combination, rather than log-linear
|
||
|
|
combination.
|
||
|
|
(This is appropriate if the input scores represent log-posterior probabilities.)
|
||
|
|
<DT><B> -non-negative </B>
|
||
|
|
<DD>
|
||
|
|
Constrain search to non-negative weight values.
|
||
|
|
<DT><B>-vocab</B><I> file</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Read the N-best list vocabulary from
|
||
|
|
<I>file</I>.<I></I><I></I><I></I>
|
||
|
|
This option is mostly redundant since words found in the N-best input
|
||
|
|
are implicitly added to the vocabulary.
|
||
|
|
<DT><B> -tolower </B>
|
||
|
|
<DD>
|
||
|
|
Map vocabulary to lowercase, eliminating case distinctions.
|
||
|
|
<DT><B> -multiwords </B>
|
||
|
|
<DD>
|
||
|
|
Split multiwords (words joined by '_') into their components when reading
|
||
|
|
N-best lists.
|
||
|
|
<DT><B>-multi-char</B><I> C</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Character used to delimit component words in multiwords
|
||
|
|
(an underscore character by default).
|
||
|
|
<DT><B> -no-reorder </B>
|
||
|
|
<DD>
|
||
|
|
Do not reorder the hypotheses for alignment, and start the alignment with
|
||
|
|
the reference words.
|
||
|
|
The default is to first align hypotheses by order of decreasing scores
|
||
|
|
(according to the initial score weighting) and then the reference,
|
||
|
|
which is more compatible with how
|
||
|
|
<A HREF="nbest-lattice.1.html">nbest-lattice(1)</A>
|
||
|
|
operates.
|
||
|
|
<DT><B>-noise</B><I> noise-tag</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Designate
|
||
|
|
<I> noise-tag </I>
|
||
|
|
as a vocabulary item that is to be ignored in aligning hypotheses with
|
||
|
|
each other (the same as the -pau- word).
|
||
|
|
This is typically used to identify a noise marker.
|
||
|
|
<DT><B>-noise-vocab</B><I> file</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Read several noise tags from
|
||
|
|
<I>file</I>,<I></I><I></I><I></I>
|
||
|
|
instead of, or in addition to, the single noise tag specified by
|
||
|
|
<B>-noise</B>.<B></B><B></B><B></B>
|
||
|
|
<DT><B>-hidden-vocab</B> file<B></B><B></B><B></B>
|
||
|
|
<DD>
|
||
|
|
Read a subvocabulary from
|
||
|
|
<I> file </I>
|
||
|
|
and constrain word alignments to only group those words that are either all
|
||
|
|
in or outside the subvocabulary.
|
||
|
|
This may be used to keep ``hidden event'' tags from aligning with
|
||
|
|
regular words.
|
||
|
|
<DT><B>-dictionary</B> file<B></B><B></B><B></B>
|
||
|
|
<DD>
|
||
|
|
Use word pronunciations listed in
|
||
|
|
<I> file </I>
|
||
|
|
to construct word alignments when building word meshes.
|
||
|
|
This will use an alignment cost function that reflects the number of
|
||
|
|
inserted/deleted/substituted phones, rather than words.
|
||
|
|
The dictionary
|
||
|
|
<I> file </I>
|
||
|
|
should contain one pronunciation per line, each naming a word in the first
|
||
|
|
field, followed by a string of phone symbols.
|
||
|
|
<DT><B>-distances</B> file<B></B><B></B><B></B>
|
||
|
|
<DD>
|
||
|
|
Use the word distance matrix in
|
||
|
|
<I> file </I>
|
||
|
|
as a cost function for word alignments.
|
||
|
|
Each line in
|
||
|
|
<I> file </I>
|
||
|
|
defines a row of the distance matrix.
|
||
|
|
The first field contains the word that is the row index,
|
||
|
|
followed by one or more word/number pairs, where the word represents the
|
||
|
|
column index and the number the distance value.
|
||
|
|
<DT><B>-init-lambdas</B><I> 'w1 w2 ...'</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Initialize the score weights to the values specified
|
||
|
|
(zeros are filled in for missing values).
|
||
|
|
The default is to set the initial acoustic model weight to 1,
|
||
|
|
the language model weight from
|
||
|
|
<B>-rescore-lmw</B>,<B></B><B></B><B></B>
|
||
|
|
the word transition weight from
|
||
|
|
<B>-rescore-wtw</B>,<B></B><B></B><B></B>
|
||
|
|
and all remaining weights to zero initially.
|
||
|
|
Prefixing a value with an equal sign (`=')
|
||
|
|
holds the value constant during optimization.
|
||
|
|
(All values should be enclosed in quotes to form a single command-line
|
||
|
|
argument.)
|
||
|
|
<BR>
|
||
|
|
Hypotheses are aligned using the initial weights; thus, it makes sense
|
||
|
|
to reoptimize with initial weights from a previous optimization in order
|
||
|
|
to obtain alignments closer to the optimimum.
|
||
|
|
<DT><B>-alpha</B><I> a</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Controls the error function smoothness;
|
||
|
|
the sigmoid slope parameter is set to
|
||
|
|
<I>a</I>.<I></I><I></I><I></I>
|
||
|
|
<DT><B>-epsilon</B><I> e</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
The step-size used in gradient descent (the multiple of the gradient vector).
|
||
|
|
<DT><B>-min-loss</B><I> x</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Sets the loss function for a sample effectively to zero when its value falls
|
||
|
|
below
|
||
|
|
<I>x</I>.<I></I><I></I><I></I>
|
||
|
|
<DT><B>-max-delta</B><I> d</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Ignores the contribution of a sample to the gradient if the derivative
|
||
|
|
exceeds
|
||
|
|
<I>d</I>.<I></I><I></I><I></I>
|
||
|
|
This helps avoid numerical problems.
|
||
|
|
<DT><B>-maxiters</B><I> m</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Stops optimization after
|
||
|
|
<I> m </I>
|
||
|
|
iterations.
|
||
|
|
In Amoeba search, this limits the total number of points in the parameter space
|
||
|
|
that are evaluated.
|
||
|
|
<DT><B>-max-bad-iters</B> n<B></B><B></B><B></B>
|
||
|
|
<DD>
|
||
|
|
Stops optimization after
|
||
|
|
<I> n </I>
|
||
|
|
iterations during which the actual (non-smoothed) error has not decreased.
|
||
|
|
<DT><B>-max-amoeba-restarts</B> r<B></B><B></B><B></B>
|
||
|
|
<DD>
|
||
|
|
Perform only up to
|
||
|
|
<I> r </I>
|
||
|
|
repeated Amoeba searches.
|
||
|
|
The default is to search until
|
||
|
|
<I> D </I>
|
||
|
|
searches give the same results, where
|
||
|
|
<I> D </I>
|
||
|
|
is the dimensionality of the problem.
|
||
|
|
<DT><B>-max-time</B><I> T</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Abort search if new lower-error point isn't found in
|
||
|
|
<I> T </I>
|
||
|
|
seconds.
|
||
|
|
<DT><B>-epsilon-stepdown</B><I> s</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
<DT><B>-min-epsilon</B><I> m</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
If
|
||
|
|
<I> s </I>
|
||
|
|
is a value greater than zero, the learning rate will be multiplied by
|
||
|
|
<I> s </I>
|
||
|
|
every time the error does not decrease after a number of iterations
|
||
|
|
specified by
|
||
|
|
<B>-max-bad-iters</B>.<B></B><B></B><B></B>
|
||
|
|
Training stops when the learning rate falls below
|
||
|
|
<I> m </I>
|
||
|
|
in this manner.
|
||
|
|
<DT><B>-converge</B><I> x</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Stops optimization when the (smoothed) loss function changes relatively by less
|
||
|
|
than
|
||
|
|
<I> x </I>
|
||
|
|
from one iteration to the next.
|
||
|
|
<DT><B> -quickprop </B>
|
||
|
|
<DD>
|
||
|
|
Use the approximate second-order method known as "QuickProp" (Fahlman 1989).
|
||
|
|
<DT><B>-init-amoeba-simplex</B><I> 's1 s2 ...'</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Perform Amoeba simplex search.
|
||
|
|
The argument defines the step size for the initial Amoeba simplex.
|
||
|
|
One value for each non-fixed search dimension should be specified,
|
||
|
|
plus optionally a value for the posterior scaling parameter
|
||
|
|
(which is searched as an added dimension).
|
||
|
|
<DT><B>-init-powell-range</B><I> 'a1</I><B>,</B><I>b1 a2</I><B>,</B><I>b2 ...'</I><B></B>
|
||
|
|
<DD>
|
||
|
|
Perform Powell search.
|
||
|
|
The argment initializes the weight ranges for Powell search.
|
||
|
|
One comma-separated pair of values for each search dimension should
|
||
|
|
be specified. For each dimension, if the upper bound equals lower bound
|
||
|
|
and initial lambda, that dimension will be fixed, even if not so specified by
|
||
|
|
<B> -init-lambda . </B>
|
||
|
|
<DT><B>-num-powell-runs</B><I> N</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Sets the number of random runs for quick Powell grid search
|
||
|
|
(default value is 20).
|
||
|
|
<DT><B> -dynamic-random-series </B>
|
||
|
|
<DD>
|
||
|
|
Use time and process ID to initialize seed for pseudo random series used
|
||
|
|
in Powell search.
|
||
|
|
This will make results unrepeatable but may yield better results through
|
||
|
|
multiple trials.
|
||
|
|
<DT><B>-print-hyps</B><I> file</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Write the best word hypotheses to
|
||
|
|
<I> file </I>
|
||
|
|
after optimization.
|
||
|
|
<DT><B>-print-top-n</B><I> N</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Write out the top
|
||
|
|
<I> N </I>
|
||
|
|
rescored hypotheses.
|
||
|
|
In this case
|
||
|
|
<B> -print-hyps </B>
|
||
|
|
specifies a directory (not a file)
|
||
|
|
and one file per N-best list is generated.
|
||
|
|
<DT><B> -print-unique-hyps </B>
|
||
|
|
<DD>
|
||
|
|
Eliminate duplicate hypotheses when writing out N-best hypotheses.
|
||
|
|
<DT><B> -print-old-ranks </B>
|
||
|
|
<DD>
|
||
|
|
Output the original hypothesis ranks when writing out N-best hypotheses.
|
||
|
|
<DT><B> -compute-oracle </B>
|
||
|
|
<DD>
|
||
|
|
Find the lowest error rate or the highest BLEU score achievable by choosing
|
||
|
|
among all N-best hypotheses.
|
||
|
|
<DT><B>-print-oracle-hyps</B><I> file</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Print output oracle hyps to
|
||
|
|
<I>file</I>.<I></I><I></I><I></I>
|
||
|
|
<DT><B>-write-rover-control</B><I> file</I><B></B><I></I><B></B><I></I><B></B>
|
||
|
|
<DD>
|
||
|
|
Writes a control file for
|
||
|
|
<B> nbest-rover </B>
|
||
|
|
to
|
||
|
|
<I>file</I>,<I></I><I></I><I></I>
|
||
|
|
reflecting the names of the input directories and the optimized parameter
|
||
|
|
values.
|
||
|
|
The format of
|
||
|
|
<I> file </I>
|
||
|
|
is described in
|
||
|
|
<A HREF="nbest-scripts.1.html">nbest-scripts(1)</A>.
|
||
|
|
The file is rewritten for each new minimal error weight combination found.
|
||
|
|
<BR>
|
||
|
|
In BLEU optimization, the weight for the ac-replacement score will be written
|
||
|
|
in the place of the posterior scale,
|
||
|
|
since posterior scaling is not used in BLEU optimization.
|
||
|
|
<DT><B> -skipopt </B>
|
||
|
|
<DD>
|
||
|
|
Skip optimization altogether, such as when only the
|
||
|
|
<B> -print-hyps </B>
|
||
|
|
function is to be exercised.
|
||
|
|
<DT><B> -- </B>
|
||
|
|
<DD>
|
||
|
|
Signals the end of options, such that following command-line arguments are
|
||
|
|
interpreted as additional scorefiles even if they start with `-'.
|
||
|
|
<DT><I>scoredir</I> ...<I></I><I></I><I></I>
|
||
|
|
<DD>
|
||
|
|
Any additional arguments name directories containing further score files.
|
||
|
|
In each directory, there must exist one file named after the sentence
|
||
|
|
ID it corresponds to (the file may also end in ``.gz'' and contain compressed
|
||
|
|
data).
|
||
|
|
The total number of score dimensions is thus 3 (for the standard scores from
|
||
|
|
the N-best list) plus the number of additional score directories specified.
|
||
|
|
</DD>
|
||
|
|
</DL>
|
||
|
|
<H2> SEE ALSO </H2>
|
||
|
|
<A HREF="nbest-lattice.1.html">nbest-lattice(1)</A>, <A HREF="nbest-scripts.1.html">nbest-scripts(1)</A>, <A HREF="nbest-format.5.html">nbest-format(5)</A>.
|
||
|
|
<BR>
|
||
|
|
S. Katagiri, C.H. Lee, & B.-H. Juang, "A Generalized Probabilistic Descent
|
||
|
|
Method", in
|
||
|
|
<I>Proceedings of the Acoustical Society of Japan, Fall Meeting</I>,
|
||
|
|
pp. 141-142, 1990.
|
||
|
|
<BR>
|
||
|
|
S. E. Fahlman, "Faster-Learning Variations on Back-Propagation: An
|
||
|
|
Empirical Study", in D. Touretzky, G. Hinton, & T. Sejnowski (eds.),
|
||
|
|
<I>Proceedings of the 1988 Connectionist Models Summer School</I>, pp. 38-51,
|
||
|
|
Morgan Kaufmann, 1989.
|
||
|
|
<BR>
|
||
|
|
W. H. Press, B. P. Flannery, S. A. Teukolsky, & W. T. Vetterling,
|
||
|
|
<I>Numerical Recipes in C: The Art of Scientific Computing</I>,
|
||
|
|
Cambridge University Press, 1988.
|
||
|
|
<BR>
|
||
|
|
<H2> BUGS </H2>
|
||
|
|
Gradient-based optimization is not supported (yet) in 1-best or BLEU mode
|
||
|
|
or in conjunction with the
|
||
|
|
<B> -combine-linear </B>
|
||
|
|
or
|
||
|
|
<B> -non-negative </B>
|
||
|
|
options;
|
||
|
|
use simplex or Powell search instead.
|
||
|
|
<BR>
|
||
|
|
The N-best directory in the control file output by
|
||
|
|
<B> -write-rover-control </B>
|
||
|
|
is inferred from the
|
||
|
|
first N-best filename specified with
|
||
|
|
<B>-nbest-files</B>,<B></B><B></B><B></B>
|
||
|
|
and will therefore only work if all N-best lists are placed in the same
|
||
|
|
directory.
|
||
|
|
<P>
|
||
|
|
The
|
||
|
|
<B> -insertion-weight </B>
|
||
|
|
and
|
||
|
|
<B> -word-weights </B>
|
||
|
|
options only affect the word error computation, not the construction
|
||
|
|
of hypothesis alignments.
|
||
|
|
Also, they only apply to sausage-based, not 1-best error optimization.
|
||
|
|
(1-best errors may be explicitly specified using the
|
||
|
|
<B> -errors </B>
|
||
|
|
option).
|
||
|
|
<P>
|
||
|
|
The
|
||
|
|
<B> -anti-refs </B>
|
||
|
|
and
|
||
|
|
<B> -anti-ref-weight </B>
|
||
|
|
options do not work for sausage-based or BLEU optimization.
|
||
|
|
<H2> AUTHORS </H2>
|
||
|
|
Andreas Stolcke <andreas.stolcke@microsoft.com>,
|
||
|
|
Dimitra Vergyri <dverg@speech.sri.com>,
|
||
|
|
Jing Zheng <zj@speech.sri.com>
|
||
|
|
<BR>
|
||
|
|
Copyright (c) 2000-2012 SRI International
|
||
|
|
<BR>
|
||
|
|
Copyright (c) 2012-2016 Andreas Stolcke
|
||
|
|
<BR>
|
||
|
|
Copyright (c) 2012-2016 Microsoft Corp.
|
||
|
|
</BODY>
|
||
|
|
</HTML>
|