source: gs3-extensions/maori-lang-detection/bin/script/gen_SentenceDetection_model.sh@ 33378

#!/bin/bash

#####################################################################################################
# Following instructions at
# https://stackoverflow.com/questions/36516363/sentence-detection-with-opennlp
# But contrary to their example, don't insert empty newlines separating the sentences input file
# Each sentence just needs to be on its own line. But no empty lines between each sentence.
#####################################################################################################

# 1. Create training file of Maori sentences in the correct format
# Train this on the 2011 set (rather than 2017) set of 100k Maori language sentences
# because the 2011 one appears to have fewer accidentally incorporated English sentences


# Need to run this script from the top level folder of this extension

if [ ! -z $1 ]; then
    if [ "x$1" = "x--help" ]; then
    echo ""
    echo "Usage: $0 <mri sentences.txt>"
    echo "   Try the one in ./opennlp-corpus/leipzig/data/mri_web_2011_100K-sentences.txt"
    echo "   (Unzip mri-opennlp-corpus.tar.gz)"
    echo "   It's better than the 100k sentences from 2017, as that contains a lot of English lines."
    echo ""
    exit
    else
    infile=$1
    fi
else
    # use the file opennlp-corpus/leipzig/data/mri_web_2011_100K-sentences.txt as inputfile
    # Check the file exists at the expected location. If not, check we have the tarball containing it and untar that.
    # If the tarball doesn't exist, try to get what we want from svn if we have svn. If no svn either, bail.
    if [ ! -f opennlp-corpus/leipzig/data/mri_web_2011_100K-sentences.txt ]; then
    if [ -f mri-opennlp-corpus.tar.gz ]; then
        tar -xvzf mri-opennlp-corpus.tar.gz
    else
        svn --help > /dev/null
        if [ "x$?" != "x0" ]; then
        echo "Attempted to train the Sentence Detector on Maori sentences on missing mri_web_2011_100K-sentences.txt"
        echo "The tarball mri-opennlp-corpus.tar.gz (containing mri_web_2011_100K-sentences.txt) didn't exist either."
        echo "And couldn't get opennlp-corpus/leipzig/data/mri_web_2011_100K-sentences.txt using SVN: no SVN."
        echo "Exitting..."
        exit
        else
        # Get just what we need from svn
        # Based on https://www.apache.org/dist/opennlp/models/langdetect/1.8.3/README.txt
        # To auto-confirm repeated svn https check, see https://serverfault.com/questions/37929/how-do-you-accept-an-ssl-certificate-through-the-svn-command-line
        svn co --depth immediates --trust-server-cert --non-interactive https://svn.apache.org/repos/bigdata/opennlp/trunk opennlp-corpus
        cd opennlp-corpus
        svn up --set-depth immediates --trust-server-cert --non-interactive
        cd leipzig
        svn up --set-depth immediates --trust-server-cert --non-interactive
        cd resources/
        svn up --set-depth infinity --trust-server-cert --non-interactive 
        cd ../data
        echo "About to download the sample Maori language 100k sentences file. This may take a while (<1 minute)."
        svn up --trust-server-cert --non-interactive mri_web_2011_100K-sentences.txt
        cd ..
        # in opennlp-corpus/leipzig
        chmod u+x create_langdetect_model.sh
        cd ../..
        fi
    fi  
    fi
    infile=./opennlp-corpus/leipzig/data/mri_web_2011_100K-sentences.txt
fi

echo $infile

#tail -100 $infile

# Ensure OPENNLP_HOME is set
if [ "x$OPENNLP_HOME" = "x" ]; then
    echo "OPENNLP_HOME not set, attempting to set it to the local apache-opennlp (v1.9.1). ENSURE THIS EXISTS OR SET OPENNLP_HOME YOURSELF!"
    #if [ -d apache-opennlp-* ]; then
    cd apache-opennlp-*
    if [ "x$?" = "x0" ]; then
    export OPENNLP_HOME=`pwd`
    cd ..
    else
    echo "No OPENNLP_HOME set and could not find a subfolder 'apache-opennlp-...' to set it to."
    echo "Set OPENNLP_HOME yourself before running this script. Exitting..."
    exit
    fi
fi

mkdir -p $OPENNLP_HOME/training_data
sentences_train_file=$OPENNLP_HOME/training_data/mri-sent.train
if [ -f "$trainfile" ]; then
    echo "Removing existing $sentences_train_file"
    rm "$sentences_train_file"
fi

# Get just the sentences in the inputfile (remove the starting number followed by tab from each line)
# see https://stackoverflow.com/questions/36516363/sentence-detection-with-opennlp
# Don't add an extra newline at end of each sentence (don't insert an empty line between each sentence)


# https://stackoverflow.com/questions/7619438/bash-read-a-file-line-by-line-and-process-each-segment-as-parameters-to-other-p
# how come it removes the first number automatically?
#while read file num sentence
#do
#    echo "$num $sentence" >> $sentences_train_file
#done < "$infile"

# Don't add the extra newline at the end, creating empty lines between sentences output,
# or the SentenceDetector in step 3 will do a poor job.
# (The example at the stackoverflow link may have included newlines separating sentences for clarity.)
#cat $infile | awk -F '\t' '{ print $2 "\n" }' > $sentences_train_file
cat $infile | awk -F '\t' '{ print $2 }' > $sentences_train_file

# Why did my way below not work?
#    openNLP-lang-detect/opennlp-corpus/leipzig/data>echo "100000\tYWCA Boarding house : ĀwhinaServices and support Kei te pūmanawa o Tāmaki Makaurau a YMCA." | awk -F "\t" '{ print $2 }'
# $1 contains everything and $2 nothing. Why?
# The problem appears to be that echo behaves differently from cat and less. Ask Dr Bainbridge what makes echo different.
# Note that I tried manually inserting \t, after copying the original line with tabspacing had no effect. Still no difference.
# Note 2: echo doesn't appear to preserve copied tab spaces.
# Answer: echo doesn't treat \n as newline and \t as tab and so on, unless the -e flag is passed in:
# echo -e "100000\tYWCA Boarding house : ĀwhinaServices and support Kei te pūmanawa o Tāmaki Makaurau a YMCA." | awk -F "\t" '{ print $2 }'

# 2. Create mri sentences model from training sentences file
#$OPENNLP_HOME/bin/opennlp SentenceDetectorTrainer -model mri-sent_trained.bin -lang en -data mri-sent.train -encoding UTF-8


mkdir -p $OPENNLP_HOME/models

$OPENNLP_HOME/bin/opennlp SentenceDetectorTrainer -model $OPENNLP_HOME/models/mri-sent_trained.bin -lang en -data $sentences_train_file -encoding UTF-8


# 3. Let's try to split some Maori language text into sentences using our model created from training on sentences.
# http://opennlp.apache.org/docs/1.9.1/manual/opennlp.html#tools.cli.sentdetect
#    Usage: opennlp SentenceDetector model < sentences
echo ""
echo "Testing the model."
echo "Running the SentenceDetector with the generated mri-sent_trained.bin model on sample_mri_paragraphs.txt..."
echo "****************************"
$OPENNLP_HOME/bin/opennlp SentenceDetector $OPENNLP_HOME/models/mri-sent_trained.bin < models-trainingdata-and-sampletxts/sample_mri_paragraphs.txt
echo "****************************"
echo ""