Released November 5, 2015





Building large LMs with SRILM

The following is a tutorial for building a large language model from the English Gigaword Fifth Edition corpus LDC2011T07 using SRILM. English text is provided from seven different sources.

Step 0: Clean up the corpus

The Gigaword corpus has to be stripped of all SGML tags and tokenized. Instructions for performing those steps are not included in this documentation. A description of this process can be found in a paper called “Annotated Gigaword”.

The Joshua package ships with a script that converts all alphabetical characters to their lowercase equivalent. The script is located at $JOSHUA/scripts/lowercase.perl.

Make a directory structure as follows:

gigaword/
├── corpus/
│   ├── afp_eng/
│   │   ├── afp_eng_199405.lc.gz
│   │   ├── afp_eng_199406.lc.gz
│   │   ├── ...
│   │   └── counts/
│   ├── apw_eng/
│   │   ├── apw_eng_199411.lc.gz
│   │   ├── apw_eng_199412.lc.gz
│   │   ├── ...
│   │   └── counts/
│   ├── cna_eng/
│   │   ├── ...
│   │   └── counts/
│   ├── ltw_eng/
│   │   ├── ...
│   │   └── counts/
│   ├── nyt_eng/
│   │   ├── ...
│   │   └── counts/
│   ├── wpb_eng/
│   │   ├── ...
│   │   └── counts/
│   └── xin_eng/
│       ├── ...
│       └── counts/
└── lm/
    ├── afp_eng/
    ├── apw_eng/
    ├── cna_eng/
    ├── ltw_eng/
    ├── nyt_eng/
    ├── wpb_eng/
    └── xin_eng/

The next step will be to build smaller LMs and then interpolate them into one file.

Step 1: Count ngrams

Run the following script once from each source directory under the corpus/ directory (edit it to specify the path to the ngram-count binary as well as the number of processors):

#!/bin/sh

NGRAM_COUNT=$SRILM_SRC/bin/i686-m64/ngram-count
args=""

for source in *.gz; do
   args=$args"-sort -order 5 -text $source -write counts/$source-counts.gz "
done

echo $args | xargs --max-procs=4 -n 7 $NGRAM_COUNT

Then move each counts/ directory to the corresponding directory under lm/. Now that each ngram has been counted, we can make a language model for each of the seven sources.

Step 2: Make individual language models

SRILM includes a script, called make-big-lm, for building large language models under resource-limited environments. The manual for this script can be read online here. Since the Gigaword corpus is so large, it is convenient to use make-big-lm even in environments with many parallel processors and a lot of memory.

Initiate the following script from each of the source directories under the lm/ directory (edit it to specify the path to the make-big-lm script as well as the pruning threshold):

#!/bin/bash
set -x

CMD=$SRILM_SRC/bin/make-big-lm
PRUNE_THRESHOLD=1e-8

$CMD \
  -name gigalm `for k in counts/*.gz; do echo " \
  -read $k "; done` \
  -lm lm.gz \
  -max-per-file 100000000 \
  -order 5 \
  -kndiscount \
  -interpolate \
  -unk \
  -prune $PRUNE_THRESHOLD

The language model attributes chosen are the following:

  • N-grams up to order 5
  • Kneser-Ney smoothing
  • N-gram probability estimates at the specified order n are interpolated with lower-order estimates
  • include the unknown-word token as a regular word
  • pruning N-grams based on the specified threshold

Next, we will mix the models together into a single file.

Step 3: Mix models together

Using development text, interpolation weights can determined that give highest weight to the source language models that have the lowest perplexity on the specified development set.

Step 3-1: Determine interpolation weights

Initiate the following script from the lm/ directory (edit it to specify the path to the ngram binary as well as the path to the development text file):

#!/bin/bash
set -x

NGRAM=$SRILM_SRC/bin/i686-m64/ngram
DEV_TEXT=~mpost/expts/wmt12/runs/es-en/data/tune/tune.tok.lc.es

dirs=( afp_eng apw_eng cna_eng ltw_eng nyt_eng wpb_eng xin_eng )

for d in ${dirs[@]} ; do
  $NGRAM -debug 2 -order 5 -unk -lm $d/lm.gz -ppl $DEV_TEXT > $d/lm.ppl ;
done

compute-best-mix */lm.ppl > best-mix.ppl

Take a look at the contents of best-mix.ppl. It will contain a sequence of values in parenthesis. These are the interpolation weights of the source language models in the order specified. Copy and paste the values within the parenthesis into the script below.

Step 3-2: Combine the models

Initiate the following script from the lm/ directory (edit it to specify the path to the ngram binary as well as the interpolation weights):

#!/bin/bash
set -x

NGRAM=$SRILM_SRC/bin/i686-m64/ngram
DIRS=(   afp_eng    apw_eng     cna_eng  ltw_eng   nyt_eng  wpb_eng  xin_eng )
LAMBDAS=(0.00631272 0.000647602 0.251555 0.0134726 0.348953 0.371566 0.00749238)

$NGRAM -order 5 -unk \
  -lm      ${DIRS[0]}/lm.gz     -lambda  ${LAMBDAS[0]} \
  -mix-lm  ${DIRS[1]}/lm.gz \
  -mix-lm2 ${DIRS[2]}/lm.gz -mix-lambda2 ${LAMBDAS[2]} \
  -mix-lm3 ${DIRS[3]}/lm.gz -mix-lambda3 ${LAMBDAS[3]} \
  -mix-lm4 ${DIRS[4]}/lm.gz -mix-lambda4 ${LAMBDAS[4]} \
  -mix-lm5 ${DIRS[5]}/lm.gz -mix-lambda5 ${LAMBDAS[5]} \
  -mix-lm6 ${DIRS[6]}/lm.gz -mix-lambda6 ${LAMBDAS[6]} \
  -write-lm mixed_lm.gz

The resulting file, mixed_lm.gz is a language model based on all the text in the Gigaword corpus and with some probabilities biased to the development text specify in step 3-1. It is in the ARPA format. The optional next step converts it into KenLM format.

Step 3-3: Convert to KenLM

The KenLM format has some speed advantages over the ARPA format. Issuing the following command will write a new language model file mixed_lm-kenlm.gz that is the mixed_lm.gz language model transformed into the KenLM format.

$JOSHUA/src/joshua/decoder/ff/lm/kenlm/build_binary mixed_lm.gz mixed_lm.kenlm