Building dictionaries

Some of you have been brave enough to start to write new language pairs for Apertium. That makes me (and all of the Apertium troop) very happy and thankful, but, most importantly, makes Apertium useful to more people.

This time I want to share some lessons I have learned after building some dictionaries: the importance of frequency estimates. For the new pairs to have the best possible coverage with a minimum of effort, it is very important to add words and rules in decreasing frequency, starting with the most frequent words and phenomena.

Frequency

A plot of word frequency in Wikipedia. The plot is in log-log coordinates. x is rank of a word in the frequency table; y is the total number of the word’s occurences. Most popular words are “the”, “of” and “and”, as expected. Zipf's law corresponds to the upper linear portion of the curve, roughly following the green (1/x) line.

A person's intuition on which words are important of frequent can be very deceptive. Therefore, the best one can do is collect a lot of text (millions of words if possible) which is representative of what one wants to translate, and study the frequencies of words and phenomena. Get it from Wikipedia, or from newspaper, or write a robot that harvests it from the web.

It is quite easy to make a crude "hit parade" of words using a simple Unix command sequence (a single line)

$ cat mybigrepresentative.txt | tr ' ' '\012' | sort -f | uniq -c | sort -nr > hitparade.txt

[I took this from Unix for Poets I think]

Of course, this may be improved a lot but serves for illustration purposes.

You will find interesting properties in this list.

One is that multiplying the rank of a word by its frequency, you get a number which is pretty constant. That's called Zipf's Law.

The other one is that half of the list are "hapax legomena" (words that appear only once).

And third, with about 1000 words you may have 75% of the text covered.

So use lists like these when you are building dictionaries.

If one of your language is English, there are interesting lists:

• Ogden's Basic English (850 words)
• Voice of America's Special English

But bear in mind that these lists are also based on a particular usage model of English, which is not "natural occurring" English.

The same applies for other linguistic phenomena. Linguists tend to focus on very infrequent phenomena which are key to the identity of a language, or on what is different between languages. But these "jewels" are usually not the "building blocks" you would use to build translation rules. So do not get carried away. Trust only frequencies and lots of real text...

Wikipedia dumps

• http://download.wikimedia.org/backup-index.html

For help in processing them see:

• http://meta.wikimedia.org/wiki/Help:Export

The dumps need cleaning up (removing Wiki syntax and XML etc.), but can provide a substantial amount of text for both frequency analysis, and sentences for POS tagger training. It can take some work, and isn't as easy as getting a nice corpus, but on the other hand they're available in ~270 languages.

You'll want the one entitled "Articles, templates, image descriptions, and primary meta-pages. -- This contains current versions of article content, and is the archive most mirror sites will probably want."

Something like (for Afrikaans):

$ bzcat afwiki-20070508-pages-articles.xml.bz2 | grep '^[A-Z]' | sed
's/$/\n/g' | sed 's/\[\[.*|//g' | sed 's/\]\]//g' | sed 's/\[\[//g' |
sed 's/&.*;/ /g'

Will give you approximately useful lists of one sentence per line (stripping out most of the extraneous formatting).

Try something like (for Afrikaans):

$ bzcat afwiki-20070508-pages-articles.xml.bz2 | grep '^[A-Z]' | sed 's/$/\n/g' | 
sed 's/\[\[.*|//g' | sed 's/\]\]//g' | sed 's/\[\[//g' | sed 's/&.*;/ /g' | tr ' ' '\012' | 
sort -f | uniq -c | sort -nr > hitparade.txt

Once you have this 'hitparade' of words, it is first probably best to skim off the top 20—30,000. Into a separate file.

$ cat hitparade.txt | head -20000 > top.lista.20000.txt

Now, if you already have been working on a dictionary then the chances are that there will exist in this 'top list' words you have already added. You can remove word forms you are already able to analyse using (for example Afrikaans):

$ cat top.lista.20000.txt | lt-proc en-af.automorf.bin  | grep '\/\*' > words_to_be_added.txt

Every 10 words or so you add, its probably worth going back and repeating this step, especially for highly inflected languages — as one lemma can produce many word forms and the wordlist is not lemmatised.

Getting cheap bilingual dictionary entries

A cheap way of getting bilingual dictionary entries between a pair of languages is as follows:

First grab yourself a wordlist of nouns in language x, for example, grab them out of the Apertium dictionary you are using:

$ cat <monolingual dictionary> | grep '<i>' | grep '__n\"' | awk -F'"' '{print $2}' 

Next, write a basic script, something like:

#!/bin/sh

#language to translate from
LANGF=$2 
#language to translate to
LANGT=$3
#filename of wordlist
LIST=$1

for LWORD in `cat $LIST`; do 
        TEXT=`wget -q http://$LANGF.wikipedia.org/wiki/$LWORD -O - | grep 'interwiki-'$LANGT`; 
        if [ $? -eq '0' ]; then
                RWORD=`echo $TEXT |  
                cut -f4 -d'"' | cut -f5 -d'/' | 
                python -c 'import urllib, sys; print urllib.unquote(sys.stdin.read());' |
                sed 's/(\w*)//g'`;
                echo '<e><p><l>'$LWORD'<s n="n"/></l><r>'$RWORD'<s n="n"/></r></p></e>'; 
        fi;
        sleep 8;
done

Note: The "sleep 8" is so that we don't put undue strain on the Wikimedia servers

And save it as iw-word.sh, then you can use it at the command line:

$ sh iw-word.sh <wordlist> <language code from> <language code to>

e.g. to retrieve a bilingual wordlist from English to Afrikaans, use:

$ sh iw-word.sh en-af.wordlist en af

The method is of variable reliability. Reports of between 70% and 80% accuracy are common. It is best for unambiguous terms, but works ok where terms retain ambiguity through languages.

Any correspondences produced by this method must be checked by native or fluent speakers of the language pairs in question.