.\" $Id: ppl-scripts.1,v 1.3 2002/04/19 14:11:30 stolcke Exp $ .TH ppl-scripts 1 "$Date: 2002/04/19 14:11:30 $" "SRILM Tools" .SH NAME ppl-scripts, add-ppls, compare-ppls, compute-best-mix, compute-best-sentence-mix, hits-from-log, ppl-from-log, subtract-ppls \- manipulate perplexities .SH SYNOPSIS .B add-ppls .RI [ ppl-file ..] .br .B subtract-ppls .I ppl-file1 .RI [ ppl-file2 ...] .br .B ppl-from-log .RI [ ppl-file ...] .br .B hits-from-log .RI [ ppl-file ...] .br .B compare-ppls [\c .BI mindelta= D\c ] .I ppl-file1 .I ppl-file2 .br .B compute-best-mix [\c .BI lambda=' "l1 l2" .RB ... ' .BI precision= P\c ] .I ppl-file1 .RI [ ppl-file2 ...] .br .B compute-best-sentence-mix [\c .BI lambda=' "l1 l2" .RB ... ' .BI precision= P\c ] .I ppl-file1 .RI [ ppl-file2 ...] .SH DESCRIPTION These scripts process the output of the .BR ngram (1) option .B \-ppl to extract various useful information. They are particularly convenient in analyzing the performance (perplexity) of language models on specific subsets of the test data, or to compare and combine multiple models. .PP .B add-ppls takes several ppl output files and computes an aggregate perplexity and corpus statistics. Its output is suitable for subsequent manipulation by .B add-ppls or .BR subtract-ppls . .PP .B subtract-ppls similarly computes an aggregate perplexity by removing the statistics of zero or more .I ppl-file2 from those in .IR ppl-file1 . Its output is suitable for subsequent manipulation by .B add-ppls or .BR subtract-ppls . .PP .B ppl-from-log recomputes the total perplexities and statistics from individual lines in .B "ngram \-debug 2 \-ppl" output. Combined with some filtering of that output this allows computing perplexities on interesting subsets of words. .PP .B hits-from-log computes N-gram hit rates from .B "ngram \-debug 2 \-ppl" output. .PP .B compare-ppls tallies the number of words for which two language models produce the same, higher, or lower probabilities. The input files should be .B "ngram \-debug 2 \-ppl" output for the two models on the same test set. The parameter .I D is the minimum absolute difference for two log probabilities to be considered different (the default is 0). .PP .B compute-best-mix takes the output of several .B "ngram \-debug 2 \-ppl" runs on the same test set and computes the optimal interpolation weights for the corresponding models, i.e., the weights that minimize the perplexity of an interpolated model. Initial weights may be specified as .IR "l1 l2 ..." . The computation is iterative and stops when the interpolation weights change by less than .I P (default 0.001). .PP .B compute-best-sentence-mix similarly optimizes the weights for sentence-level interpolation of LMs. It requires input files generated by .BR "ngram \-debug 1 \-ppl" . (Sentence-level mixtures can be implemented using the .B "ngram \-hmm" option, by constructing a suitable HMM structure.) .SH "SEE ALSO" ngram(1). .SH BUGS All scripts depend on the idiosyncrasies of .B "ngram \-ppl" output. .SH AUTHOR Andreas Stolcke . .br Copyright 1995-2002 SRI International
