Calculating coverage

En français

Simple bidix-trimmed coverage testing

First install apertium-cleanstream:

svn checkout https://svn.code.sf.net/p/apertium/svn/trunk/apertium-tools/apertium-cleanstream
cd apertium-cleanstream
make
sudo cp apertium-cleanstream /usr/local/bin

Then save this as coverage.sh:

#!/bin/bash
mode=$1
outfile=/tmp/$mode.clean
apertium -d . $mode | apertium-cleanstream -n > $outfile
total=$(grep -c '^\^' $outfile)
unknown=$(grep -c '/\*' $outfile)
bidix_unknown=$(grep -c '/@' $outfile)
known_percent=$(calc -p  "round( 100*($total-$unknown-$bidix_unknown)/$total, 3)")
echo "$known_percent % known tokens ($unknown unknown, $bidix_unknown bidix-unknown of total $total tokens)"
echo "Top unknown words:"
grep '/[*@]' $outfile | sort | uniq -c | sort -nr | head

And run it like

cat asm.corpus | bash coverage.sh asm-eng-biltrans

(The bidix-unknown count should always be 0 if your pair uses automatic analyser trimming.)

TODO: paradigm-coverage (less naïve)

On an analysed corpus, we can sum frequencies into bins for each lemma+mainpos, so if the analysed corpus contains

musa/mus<n><f><sg><def>/muse<vblex><past>
mus/mus<n><f><sg><ind>/mus<n><f><pl><ind>/muse<vblex><imp>
musene/mus<n><f><pl><def>

then output has

3 mus<n><f>
2 muse<vblex>

and we can find paradigms that are likely to mess up disambiguation, or where we need to ensure that the bidix contains the highest-frequency paradigm (since the bidix is typically smaller than the monodix).

We could also weight these numbers by number of unique forms in the pardef; if the verb pardef has 6 unique forms and then noun only 3, then the above output should be even more skewed:

0.33 mus<n><f>
0.75 muse<vblex>

Faster coverage testing with frequency lists

If words appear several times in your corpus, why bother analysing them several times? We can make a frequency list first and add together the frequencies. This script does some very stupid tokenisation and creates a frequency list:

make-freqlist.sh:

#!/bin/bash

if [[ -t 0 ]]; then
    echo "Expecting a corpus on stdin"
    exit 2
fi

tr '[:space:][:punct:]' '\n' | grep . | sort | uniq -c | sort -nr

And this script runs your analyser, summing up the frequencies:

freqlist-coverage.sh:

#!/bin/bash

set -e -u

if [[ $# -eq 0 || -t 0 ]]; then
    echo "Expecting apertium arguments and a 'sort|uniq -c|sort -nr' style frequency list on stdin"
    echo "For example:"
    echo "\$ < spa.freqlist $0 -d . spa-morph"
    exit 2
fi

sed 's%^ *%<apertium-notrans>%;s% %</apertium-notrans>%;s%$% .%' |
apertium -f html-noent "$@" |
awk -F'</?apertium-notrans>| *\\^\\./\\.<sent><clb>\\$' '
/[/][*@]/ {
    unknown+=$2
    if(!printed) print "Top unknown tokens:"
    if(++printed<10) print $2,$3
    next
}
{
    known+=$2
}
END {
    total=known+unknown
    known_pct=100*known/total
    unk_pct=100*unknown/total
    print known_pct" % known of total "total" tokens"
}'

Usage:

$ chmod +x make-freqlist.sh freqlist-coverage.sh
$ bzcat ~/corpora/nno.txt.bz2 |./make-freqlist.sh > nno.freqlist
$ <nno.freqlist ./freqlist-coverage.sh -d ~/apertium-svn/languages/apertium-nno/ nno-morph

coverage.py

https://svn.code.sf.net/p/apertium/svn/trunk/apertium-tools/coverage.py is a coverage script that wraps curl and bzcat (?)

More involved scripts

Often it's nice to clean up wikipedia fluff etc. for coverage testing.

Notes on calculating coverage from wikipedia dumps (based on Asturian#Calculating coverage).

(Mac OS X `sed' doesn't allow \n in replacements, so I just use an actual (escaped) newline...)

wikicat.sh:

#!/bin/sh
# Clean up wikitext for running through apertium-destxt

# awk prints full lines, make sure each html element has one
bzcat "$@" | sed 's/>/>\
/g' | sed 's/</\
</g' |\
# want only stuff between <text...> and </text>
awk '
/<text.*>/,/<\/text>/ { print $0 }
' |\
sed 's/\./ /g' |\
# Drop all transwiki links
sed 's/\[\[\([a-z]\{2,3\}\|bat-smg\|be-x-old\|cbk-zam\|fiu-vro\|map-bms\|nds-nl\|roa-rup\|roa-tara\|simple\|zh-classical\|zh-min-nan\|zh-yue\):[^]]\+\]\]//g' |\
# wiki markup, retain bar and fie from [[foo|bar]] [[fie]]
sed 's/\[\[[^]|]*|//g' | sed 's/\]\]//g' | sed 's/\[\[//g' |\
# wiki markup, retain `bar fie' from [http://foo bar fie] and remove [http://foo]
sed 's/\[http[^ ]*\([^]]*\)\]/\1/g' |\
# remove entities
sed 's/&[^;]*;/ /g' |\
# and put space instead of punctuation
sed 's/[;:?,]/ /g' |\
# Keep only lines starting with a capital letter, removing tables with style info etc.
grep '^[ 	]*[A-ZÆØÅ]' # Your alphabet here

count-tokenized.sh:

#!/bin/sh
# http://wiki.apertium.org/wiki/Asturian#Calculating_coverage

# Calculate the number of tokenised words in the corpus:
apertium-destxt | lt-proc $1 |apertium-retxt |\
# for some reason putting the newline in directly doesn't work, so two seds
sed 's/\$[^^]*\^/$^/g' | sed 's/\$\^/$\
^/g' 

To find all tokens from a wiki dump:

$ ./wikicat.sh nnwiki-20090119-pages-articles.xml.bz2 > nnwiki.cleaned.txt
cat nnwiki.cleaned.txt | ./count-tokenized.sh nn-nb.automorf.bin | wc -l

To find all tokens with at least one analysis (naïve coverage):

$ cat nnwiki.cleaned.txt  | ./count-tokenized.sh nn-nb.automorf.bin | grep -v '\/\*' | wc -l

To find the top unknown tokens:

$ cat nnwiki.cleaned.txt  | ./count-tokenized.sh nn-nb.automorf.bin | sed 's/[ 	]*//g' |\ # tab or space
   grep '\/\*' | sort -f | uniq -c | sort -gr | head 

Script ready to run

corpus-stat.sh

#!/bin/sh
# http://wiki.apertium.org/wiki/Asturian#Calculating_coverage


# Example use:
# zcat corpa/en.crp.txt.gz | sh corpus-stat.sh


#CMD="cat corpa/en.crp.txt"
CMD="cat"

F=/tmp/corpus-stat-res.txt

# Calculate the number of tokenised words in the corpus:
# for some reason putting the newline in directly doesn't work, so two seds
$CMD | apertium-destxt | lt-proc en-eo.automorf.bin |apertium-retxt | sed 's/\$[^^]*\^/$^/g' | sed 's/\$\^/$\
^/g' > $F

NUMWORDS=`cat $F | wc -l`
echo "Number of tokenised words in the corpus: $NUMWORDS"



# Calculate the number of words that are not unknown

NUMKNOWNWORDS=`cat $F | grep -v '\*' | wc -l`
echo "Number of known words in the corpus: $NUMKNOWNWORDS"


# Calculate the coverage

COVERAGE=`calc "round($NUMKNOWNWORDS/$NUMWORDS*1000)/10"`
echo "Coverage: $COVERAGE %"

#If you don't have calc, change the above line to:
#COVERAGE=$(perl -e 'print int($ARGV[0]/$ARGV[1]*1000)/10;' $NUMKNOWNWORDS $NUMWORDS)

# Show the top-ten unknown words.

echo "Top unknown words in the corpus:"
cat $F | grep '\*' | sort -f | uniq -c | sort -gr | head -10

Sample output:

$ zcat corpa/en.crp.txt.gz | sh corpus-stat.sh
Number of tokenised words in the corpus: 478187
Number of known words in the corpus: 450255
Coverage: 	94.2 %
Top unknown words in the corpus:
    191 ^Apollo/*Apollo$
    104 ^Aramaic/*Aramaic$
     91 ^Alberta/*Alberta$
     81 ^de/*de$
     80 ^Abu/*Abu$
     63 ^Bakr/*Bakr$
     62 ^Agassi/*Agassi$
     59 ^Carnegie/*Carnegie$
     58 ^Agrippina/*Agrippina$
     58 ^Achilles/*Achilles$
     56 ^Adelaide/*Adelaide$

See also

• Wikipedia dumps
• apertium-cleanstream