Skip to content

Latest commit

 

History

History
88 lines (62 loc) · 3.92 KB

README.md

File metadata and controls

88 lines (62 loc) · 3.92 KB
Raw

Bitdeli Badge

SWIG-SRILM: A SWIG-based wrapper for an SRILM language model

==========

Description

This package contains files to generate Perl and Python wrappers for SRILM language models.

Requirements

  • GNU make
  • Simplified Wrapper & Interface Generator (SWIG)
  • A local Python and/or Perl installation
  • The SRILM toolkit (v1.7.1). If you have an older version of SRILM e.g., the 1.5.x series then you should use the old_srilm branch. Note that SRILM should have been compiled as position independent code. You can do that by using the command MAKE_PIC=yes make when compiling SRILM.

Installation:

Linux

  • Modify the following environment variables at the top of the included Makefile:

  • SRILM_LIBS : The directory containing the SRILM libraries

  • SRILM_INC : The directory containing the SRILM header files

  • PYTHON_INC : The directory containing the python header files

  • PERL_INC : The directory containing the perl header files

  • To create a Python module, run 'make python' in this directory. Copy _srilm.so and srilm.py to your directory where you want to use the python module. You can run the included test.py script to check whether the compiled module works correctly. The output of test.py should be the following:

1. Number of n-grams:
   There are 11868 unigrams in this LM
   There are 59481 bigrams in this LM
   There are 16744 trigrams in this LM
   There are 13787 4-grams in this LM
   There are 12082 5-grams in this LM

2. N-gram log probabilities:
   p('good') = -3.49373698235
   p('of the') = -0.558740794659
   p('nitin madnani') = -99.0
   p('there are some') = -0.985605716705
   p('do more about your') = -0.469523012638
   p('or whatever has yet to') = -0.53226429224

3. Sentence log probabilities and perplexities:
   p('there are some good') = -9.85836982727
   ppl('there are some good') = 93.6858444214

4. OOvs:
   nOOVs('there are some foobar') = 1

5. Corpus log probabilties and perplexities:
   Logprob for the file test.txt = -33.6016654968
   Perplexity for the file test.txt = 94.7476806641
  • To create a Perl module, run make perl in this directory. Copy srilm.so and srilm.pm to the directory of your choice. Run the included Perl script 'test.pl' to test whether the compiled module works correctly. The output should be the same as above.

Mac OS X

Note: This has only been tested on OS X El Capitan and only with the built-in versions of python (2.7.10) and perl (5.18).

  • Check out the macosx branch.

  • Make sure you have compiled the SRILM libraries (MAKE_PIC=yes make).

  • Go to the directory containing the SRILM header files ($SRILM/include), open File.h and comment out the line that says #include "zio.h". This is necessary because even though SRILM is supposed to rename the zopen() function to my_zopen() on OS X since zlib is installed by default, it does not seem to work. So, this is a hacky workaround.

  • Modify the following environment variables at the top of Makefile.osx:

  • SRILM_LIBS : The directory containing the SRILM libraries

  • SRILM_INC : The directory containing the SRILM header files

    IMPORTANT: DO NOT change the PYTHON_INC and PERL_INC variables as they are set to be the default values for OS X El Capitan.

  • To compile the python module, run make -f Makefile.osx python and to compile the perl module, run make -f Makefile.osx perl. Note that the compiled modules will only work with the default OS X python and perl interpreters, i.e., /usr/bin/python and /usr/bin/perl.

  • You should be able to run /usr/bin/python test.py and /usr/bin/perl test.pl to test that the modules work and obtain the same output in the Linux case.

Usage:

Usage is clearly illustrated in files test.pl and test.py.